Professional Documents
Culture Documents
Robust Adaptive MIMO Control Using Multiple-Model Hypothesis Testing and Mixed-Mu Synthesis
Robust Adaptive MIMO Control Using Multiple-Model Hypothesis Testing and Mixed-Mu Synthesis
By
Sajjad Fekri
i
“The true logic of this world is in the calculus of probabilities.”
– James C. Maxwell
ii
Abstract
The thesis addresses the problem of robust adaptive control for linear multivari-
able systems subject to unmodeled dynamics and uncertain real parameters, either
constant or slowly time-varying. The design methodology integrates state-of-the-art
non-adaptive robust feedback synthesis techniques, using the mixed-µ design method
(using the so-called D,G-K iteration), with a stochastic system identification method
that exploits multiple-models and dynamic hypothesis-testing concepts.
The proposed design methodology, Robust Multiple-Model Adaptive Control, re-
ferred to as RMMAC, leads to an efficient procedure for the design of “robust” adap-
tive multivariable feedback control systems for plants that include both unmodeled
dynamics and possibly large parametric uncertainty in the plant state description.
In the course of the design process, the minimum number of required models, their
parameter uncertainty range and also local compensators are automatically obtained
by using the proposed step-by-step adaptive methodology. The RMMAC methodol-
ogy is applicable to multivariable feedback control system designs with full accounting
of model parameter, dynamic errors, and explicit performance requirements.
The performance of the RMMAC methodology is evaluated using non-trivial sim-
ulation examples, with one or two unknown parameters in both SISO and MIMO
systems, and is compared with those obtained with the best global non-adaptive ro-
bust control (GNARC). The RMMAC can be an attractive engineering design tool
for disturbance-rejection and command-tracking and a viable alternative for control
applications where the performance that can be achieved with the GNARC is unsat-
isfactory.
Index Terms– Multivariable robust adaptive control, robust µ-synthesis, multiple-
model estimators and compensators, adaptive systems.
iii
List of Symbols and Notations
b damping coefficient A state matrix
filtered continuous plant white noise
d B input matrix
also Baram proximity measure
k spring stiffness C output matrix
mass
m D feedforward matrix
also number of measurements
iv
Table of Contents
List of Tables x
List of Figures xi
Acknowledgments xxiii
1 Introduction 1
1.1 What is the Problem? . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Significance of the Problem . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Brief Review of Related Literature . . . . . . . . . . . . . . . . . . . 4
1.4 The Contributions of This Thesis . . . . . . . . . . . . . . . . . . . . 7
1.4.1 Remark on Engineering Relevance . . . . . . . . . . . . . . . . 12
1.5 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
v
3.4.2 Information Definition and Properties . . . . . . . . . . . . . . 58
3.4.3 Application to Linear Systems . . . . . . . . . . . . . . . . . . 60
3.5 The Baram Proximity Measure (BPM) . . . . . . . . . . . . . . . . . 62
3.5.1 Stationarity and Ergodicity Properties . . . . . . . . . . . . . 68
3.5.2 Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.5.3 Convergence Conditions . . . . . . . . . . . . . . . . . . . . . 72
3.5.4 Robustifying Kalman Filters using Fake White Noise . . . . . 72
3.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
vi
5.6 RMMAC with Switching Architecture (RMMAC/S) . . . . . . . . . . 138
5.7 RMMAC/XI Architecture: Dealing with Unknown Plant White Noise
Intensity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
5.8 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
vii
7.7.1 RMMAC Design for the NMP Plant with X%=50% . . . . . . 261
7.7.2 RMMAC Design for the NMP plant with X%=60% . . . . . . 263
7.7.3 A “Brute-Force” Approach . . . . . . . . . . . . . . . . . . . . 264
7.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
viii
10.3.4 Relations between the Baram Proximity Measure (BPM) and
the Vinnicombe Metric . . . . . . . . . . . . . . . . . . . . . . 324
10.3.5 RMMAC with Three or More Uncertain Parameters . . . . . . 325
10.3.6 RMMAC with Unstable Plant . . . . . . . . . . . . . . . . . . 326
10.3.7 Linear Parameter Varying (LPV) Systems and RMMAC . . . 326
10.3.8 Designing “Stability Safety Nets” for the RMMAC . . . . . . 327
10.3.9 Global Stability Theory for MMAC Architectures . . . . . . . 328
Bibliography 329
ix
List of Tables
5.1 List of Performance Evaluation Tools . . . . . . . . . . . . . . . . . . 116
7.1 Best Ap ’s of GNARC and each LNARC for the NMP plant . . . . . . 230
7.2 The mismatched-model stability cases for the NMP plant . . . . . . . 237
7.3 LNARC stability maintenance for the NMP plant . . . . . . . . . . . 238
7.4 Best Ap ’s of LNARCs based on X%=50% . . . . . . . . . . . . . . . . 263
7.5 Best Ap ’s of LNARCs based on X%=60% . . . . . . . . . . . . . . . . 264
7.6 Best Ap ’s of LNARCs for the NMP plant based on a Brute-Force approach265
x
List of Figures
1.1 The classical MMAE and the Classical MMAC (CMMAC) architecture. 5
1.2 The proposed RMMAC methodology. . . . . . . . . . . . . . . . . . . 9
xi
2.9 RMMAC philosophy in the two-dimensional parameter case. The large
bound in the plant real uncertainties is decomposed into N small model
spaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
xii
5.11 Switching RMMAC (RMMAC/S) Architecture. . . . . . . . . . . . . 139
5.12 The RMMAC/XI architecture. . . . . . . . . . . . . . . . . . . . . . . 142
6.1 The two-cart system. The spring stiffness, k1 , is uncertain, and there
is an unmodeled time-delay in the control channel. . . . . . . . . . . . 146
6.2 Bode plot of the open-loop MSD system for different values of the
spring constant k1 , eq. (6.5). . . . . . . . . . . . . . . . . . . . . . . . 149
6.3 Frequency weight of unmodeled time-delay dynamic used in the GNARC
and each LNARC designs and its frequency bound, for τ = 0.05 secs. 150
6.4 Frequency-weighted functions used in the LFD. . . . . . . . . . . . . 152
6.5 MSD system with weights for the mixed-µ synthesis. . . . . . . . . . 153
6.6 Best GNARC and FNARC performance gains for spring-stiffness un-
certainty as in eq. (6.5). Note the logarithmic scale. . . . . . . . . . . 154
6.7 Best local performance gains using mixed-µ using 70% of the best
FNARC performance designed for local uncertainty bounds. . . . . . 155
6.8 Frequency-domain characteristics (Bode plot) of the compensators GNARC
and the four LNARCs. . . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.9 The four-model RMMAC architecture for the MSD example (LFD case).158
6.10 Potential improvement from RMMAC visualized by the Bode plots of
disturbance-rejection transfer function |Mdz (jω)|. . . . . . . . . . . . 159
6.11 Bode plot of transfer functions from ξ(t) or θ(t) to control u(t). . . . 160
6.12 Predicted potential RMS performance of the RMMAC vs GNARC,
with no sensor noise, i.e. θ(t) = 0. . . . . . . . . . . . . . . . . . . . . 161
6.13 Predicted potential RMS performance of the RMMAC vs GNARC with
both ξ and θ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
6.14 The mismatch-instability cases where each local compensator Kj (s),
LNARC#j, is tested to be placed in the feedback while the uncertain
parameter (spring constant k1 ) varies over all possible values in the set,
Ω.
×: shows an unstable model/compensator pair, and ¤: shows a stable
pair. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
xiii
6.15 Baram Proximity Measures (BPMs) for the optimized four KFs (LFD
case). The ♦ show the nominal spring constant used to design the
associated KF, see eq. (6.22). . . . . . . . . . . . . . . . . . . . . . . 164
6.16 “Easy” identification I: Simulation results for k1 = 1.65 (LFD case). . 167
6.17 “Easy” identification II: Simulation results for k1 = 0.3 (LFD case). . 168
6.24 The RMMAC evaluation under severe assumption violation. The plant
disturbance intensity is unit (Ξ = 1) in designing the KFs, but in the
simulations it is chosen Ξact = 100. The value of the spring constant is
k1 = 1.5. Only one MC run is shown. Note that the RMMAC exhibits
poor performance in comparison to the GNARC, due to the frequent
switching of the posterior probabilities (LFD case). . . . . . . . . . . 178
xiv
6.26 The RMMAC performance evaluation under temporary severe viola-
tion. The designing plant disturbance intensity is unit (Ξd = 1) but
it is a periodic square-wave in the simulations (Ξact ∈ {1, 100}). The
value of the spring stiffness is k1 = 0.9. Only one MC run is shown
(LFD case). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
6.27 The RMMAC evaluation under long-term forced instability (48.5 ≤ t ≤
51.5) for k1 = 1.5 in Model#1. Only one MC run is shown. However,
the probability of such an event due to “abnormal” plant disturbance
and sensor noise signals is truly infinitesimal (LFD case). Note that
the RMMAC recovers after 170 secs. . . . . . . . . . . . . . . . . . . 181
6.28 RMMAC responses for the fast sinusoidal parameter variations, as in
eq. (6.27) (LFD case). . . . . . . . . . . . . . . . . . . . . . . . . . . 183
6.29 The RMMAC/XI architecture for the MSD example (LFD case). . . . 185
6.30 The RMMAC/XI performance for varying plant noise intensity. The
spring constant k1 = 0.9 is in Model#2 with Ξ = 1 and in Model#6
with Ξ = 100. Note that the correct probabilities, P2 (t) and P6 (t),
(almost immediately) go to unity as Ξ changes (LFD case). . . . . . . 186
6.31 Performance of the RMMAC and Switching RMMAC (RMMAC/S)
(LFD case). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
6.32 Bode plot of the open-loop MSD system for different values of spring
constant k1 (HFD case). . . . . . . . . . . . . . . . . . . . . . . . . . 188
6.33 Frequency-weighted functions used in the HFD. . . . . . . . . . . . . 189
6.34 Best GNARC and FNARC performance parameters for the HFD cal-
culated for all k1 ∈ Ω = [0.25, 1.75]. . . . . . . . . . . . . . . . . . . . 190
6.35 Best possible HFD performance parameters, using DGK algorithm,
requiring 66% of the best FNARC performance. . . . . . . . . . . . . 190
6.36 Frequency-domain characteristics of the GNARC and LNARCs com-
pensators for the HFD. . . . . . . . . . . . . . . . . . . . . . . . . . . 192
6.37 RMMAC potential improvement via plots of the disturbance-rejection
transfer function |Mdz (jω)| (HFD case). . . . . . . . . . . . . . . . . 193
xv
6.38 Predicted potential RMS performance of the RMMAC vs GNARC due
to xi(t) only (HFD case). (a) Output RMS comparisons, (b) Control
RMS comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
6.39 BPM for the HFD. The nominal k1 of the KFs (not shown) are at the
minimum values of the Li (BPM) curves; see eq. (6.34). . . . . . . . . 195
6.40 Monte Carlo stochastic simulation results when true spring constant is
k1 = 1.65 (HFD case). . . . . . . . . . . . . . . . . . . . . . . . . . . 197
6.41 Monte Carlo stochastic simulation results when true spring constant is
k1 = 0.45 (HFD case). . . . . . . . . . . . . . . . . . . . . . . . . . . 197
6.45 Simulation results of the RMMAC performance when the actual distur-
bance d(t) is a ±2 periodic square-wave with the period of T = 60 secs,
and k1 = 1.75 (HFD case). Only one MC run is shown. . . . . . . . . 202
6.46 Monte Carlo simulation results of the RMMAC performance for sinu-
soidal sensor noise when k1 = 1.75 (HFD case). . . . . . . . . . . . . 203
6.47 The RMMAC performance evaluation with slow sinusoidal k1 (t) vari-
ation (HFD case). Only one MC run is shown. . . . . . . . . . . . . . 204
6.48 Case 1A. The RMMAC evaluation under severe assumption violations.
The plant disturbance intensity is unit (Ξ = 1) when designing the
KFs, but in the simulations, we use Ξact = 100 (HFD case). The value
of the spring constant is k1 = 1.65. Only one MC run is shown. . . . 206
xvi
6.49 Case 1B. The RMMAC evaluation under severe assumption violations.
The plant disturbance intensity is Ξ = 1 when designing the KFs, but
in the simulations it switches in Ξact ∈ {1, 100} (HFD case). The value
of the spring constant is k1 = 0.3. Only one MC run is shown. . . . . 207
6.50 Case 1C. The RMMAC performance evaluation under severe assump-
tion violation where the designing plant disturbance intensity is Ξ = 1
but it is very small in the simulations, Ξact = 0.01 (HFD case). The
value of the spring stiffness is k1 = 0.4 in Model#6. Only one MC run
is shown. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
6.51 Forced mismatch instability for 50 ≤ t ≤ 65 secs (HFD case). The true
value is k1 = 0.6 in Model#5. Note the rapid recovery of the RMMAC
after the forced instability, i.e. P5 (t) → 1. Only one MC is shown. . . 210
6.52 RMMAC evaluation with fast sinusoidal parameter variation, as in
eq. (6.37) (HFD case). . . . . . . . . . . . . . . . . . . . . . . . . . . 211
6.53 The RMMAC/XI performance for the HFD when Ξ is a periodic “square-
wave” as in 6.53(a). The true spring constant is k1 = 0.3 in Model#7
when Ξ = 1 and in Model#14 when Ξ = 100. Note that the “correct”
posterior probabilities P7 (t) and P14 (t) quickly respond to the changes
in Ξ in 6.53(a). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
6.54 The RMMAC/XI performance compared to the standard RMMAC
subject to the unknown plant noise intensities. The true spring is
k1 = 0.3 (HFD case). . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
6.55 Performance evaluation of the RMMAC and Switching RMMAC/S.
Only one MC run is shown. The true value of spring constant is k1 =
0.5 in Model#5, but also close to Model#6 (HFD case). . . . . . . . 215
6.56 Comparison of performance gains for the MSD system with one and
two measurements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
xvii
7.5 Frequency weights used in the GNARC and each LNARC design. . . 226
7.6 Best GNARC and FNARC performance gains for the NMP plant. Note
that the scale of both the Ap -axis and the Z-axis is logarithmic. . . . 227
7.7 “Best” performance gains for each of the four LNARCs using the pro-
posed RMMAC algorithm. The LNARCs are designed for the uncer-
tainty subsets by considering X% = 30% of the best FNARC perfor-
mance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
7.9 Baram proximity measures (BPMs) based on the nominal NMP zeros
given in eq. (7.12) used in the KF designs. . . . . . . . . . . . . . . . 232
7.10 The four-model RMMAC architecture for the NMP example. . . . . . 233
7.15 Case 1A. Performance characteristics, the output z(t) vs time due to
stochastic inputs. The true value of the NMP zero is Z = 4.89. . . . . 240
7.17 Case 1B. Performance characteristics of the GNARC and the RMMAC,
the output z(t) vs time due to stochastic inputs. The true value of the
NMP zero is Z = 100.0 . . . . . . . . . . . . . . . . . . . . . . . . . . 241
xviii
7.18 The behavior of 10 Monte Carlo samples for the posterior probabilities
for Case 2A. The true value of the NMP zero is Z = 1.0. Note that
P1 (t) rapidly decays from its initial value P1 (0) = 0.97 to zero in less
than a second. The numerical average is also shown (solid curves). . . 244
7.20 The behavior of 10 Monte Carlo samples for the posterior probabili-
ties for Case 2B. Note that P3 (t) rapidly decays from its forced value
P3 (0) = 0.97 to zero in less than a second. The numerical average is
also shown (solid curves). The true value of zero Z = 1. . . . . . . . . 246
7.21 The performance of the RMMAC for Case 2B, averaging 10 Monte
Carlo runs, when we enforce temporary instability at t = 30 secs. The
true value of zero Z = 1. . . . . . . . . . . . . . . . . . . . . . . . . . 247
7.22 The transient behavior of the posterior probabilities. The plant distur-
bance includes a deterministic sinusoidal component plus filtered white
noise, eq. (7.17). Z = 10. . . . . . . . . . . . . . . . . . . . . . . . . . 248
7.23 Transient responses of the output and control signals when the plant
disturbance is a sinusoid plus filtered white noise and Z = 10. . . . . 249
7.25 The RMMAC evaluation under mild assumption violating with sinu-
soidal sensor noise with Z = 20. An average of 10 MC runs are shown. 252
7.26 RMMAC responses for the sinusoidal parameter variations, as in eq. (7.20).254
xix
7.29 Case 3A. The RMMAC evaluation under severe assumption violation.
The plant disturbance intensity is unit (Ξ = 1) in designing the KFs,
but in the simulations a very large disturbance, with Ξact = 100, drives
the NMP plant. The value of the NMP zero is Z = 50. Only one MC
run is shown. Note the poor performance of the RMMAC. . . . . . . 259
7.30 The RMMAC performance evaluation under severe assumption viola-
tion where the designing plant disturbance intensity is unit (Ξ = 1)
but a very small, Ξact = 0.01, is used in the simulations. The value of
the NMP zero is Z = 1.50 and only one MC run is shown. . . . . . . 261
7.31 Best performance gains for each of the six LNARCs by requiring X% =
50% of the best FNARC performance curve. . . . . . . . . . . . . . . 262
7.32 Best performance gains for each of the eight LNARCs by requiring
X% = 60% of the best FNARC performance curve. . . . . . . . . . . 263
7.33 LNARCs for the NMP plant by using a Brute-Force approach. . . . . 265
9.1 The TITO three-cart MSD system. The spring k1 and the mass m3
are uncertain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
9.2 The singular value Bode plots of the TITO MSD system. . . . . . . . 289
9.3 MIMO weights for Robust Synthesis. Control accuracy of z2 (t) is se-
lected to be four times more important than z1 (t). . . . . . . . . . . . 292
xx
9.4 The five-model RMMAC architecture for this example. . . . . . . . . 293
9.5 The best performance parameters, AFp ’s, of the FNARC found at the
corners of the 2-D uncertain rectangle. . . . . . . . . . . . . . . . . . 294
9.10 BPM#1 and BPM#2, with the fake noise intensity of eq. (9.26). . . . 302
9.11 Design points for each MIMO KF selected (not at the model center)
so that the BPMs for each adjacent model boundary are matched. . . 302
xxi
9.15 Simulation results of the RMMAC performance when the actual distur-
bances d(t) are sinusoidal disturbances, with period of 60 secs, super-
imposed with the original stochastic disturbances. The values k1 = 1.2
and m3 = 1.7 in Model#2 (and also close to Model#1) are used. Only
one MC run is shown. . . . . . . . . . . . . . . . . . . . . . . . . . . 308
9.16 RMMAC recovering from forced instability at t = 30 secs. The values
k1 = 1.5, m3 = 1.0 (in M#1) is used. . . . . . . . . . . . . . . . . . . 310
9.17 RMMAC with time-varying parameters. Simulation time (4000 s) is
slightly longer than one complete sweep period (3769 s), i.e. a complete
cycle is shown in (a). . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
9.18 The RMMAC behavior for varying plant noises intensity. The spring
constant k1 = 0.3 and the mass m3 = 0.3 are in Model#5. The KFs
were designed using Ξ = 1. . . . . . . . . . . . . . . . . . . . . . . . 312
9.19 Comparison of the standard RMMAC and the RMMAC performance
for varying plant noises intensity as in Figure 9.18(a). The spring
constant k1 = 0.3 and the mass m3 = 0.3 are in Model#5 with Ξ = 1
and in Model#10 with Ξ = 100. Note that the correct probabilities,
P5 (t) and P10 (t), (almost immediately) go to unity when Ξ changes. . 314
xxii
Acknowledgments
This is a good opportunity to acknowledge the help and support of a number of
friends, colleagues, and instructors, who have helped me during the period of my PhD
studies in Lisbon. Without their help and support this dissertation would never have
evolved to its present form. Here is my attempt to express my deepest appreciation
to everyone.
First of all, I would like to express my gratitude to my research supervisor, Profes-
sor Michael Athans, for providing the initial motivation to my research work, for his
spirit of innovation, for sticking with me, and for many of the ideas discussed in this
dissertation. I appreciate his leadership and support even though I know I made his
life stressful at times. I will always cherish the stimulating, sometimes painstaking
discussions that we had over the past two years, during our almost daily meetings,
often during the weekends. As an advisor, he made me enthusiastic about research,
while keeping me on track, and inspired me to better understand what questions a
control system designer must answer. He is an intelligent (and knowledgeable) sci-
entist and engineer. As a friend, he has outstanding human qualities; my career has
been totally influenced by him. I am forever grateful to him for his past, present, and
future support.
I would also like to thank my thesis supervisor, Professor Antonio Pascoal, for his
timely advice and insightful discussions, and for following my progress throughout
this thesis. He is inspiring in his ability to actively seek new projects and pursue
them creatively. Besides, I greatly profited from his efforts to navigate through all
bureaucracies encountered in any graduate student’s life.
Foremost among those who helped me during my stay at ISR/IST is my wife,
Shabnam. I would like to thank her for all her endless love, patience, understanding,
and support in keeping me focused on my goals. Without her unconditional support,
this dissertation would have taken much longer to be completed. I am also indebted
xxiii
to my family, most importantly my parents, Jalal and Oranous, and my wife’s family,
due to their encouragement and help.
I would also like to thank the members of my doctoral committee, Prof. João
Lemos and Prof. Fernando Pereira, for their generosity and enthusiasm about re-
search. My sincere thanks go also to Prof. Joao Sentieiro, director of ISR, for his
support since I was accepted at the Institute.
I also wish to thank Prof. Gary Balas, for the mixed-µ synthesis software, Prof.
Yoram Baram for a copy of his MIT Ph.D. thesis, and Prof. Peter Young and Dr.
Chris Greene for providing papers that were instrumental to my research.
Also, among those who deserve credit are Prof. Brian Anderson and Prof. João
Hespanha for many stimulating comments, suggestions, and criticisms; their astute
technical comments gave me a real insight into understanding and smoothing out some
parts of this dissertation. I would also like to thank Profs. Rita Almeida Ribeiro,
from University of Nova Lisboa (Uninova), Isabel Ribeiro, from IST, and Stephanie
Gettelfinger for their help and support in our personal affairs.
Finally, I appreciate the help and friendship of other people at ISR, especially, José
Alberto Victor, Luis Sebastiao, Reza Ghabcheloo, Manuel Rufino, Pedro Gomes, José
Vasconcelos, João Xavier, João Pedro Gomes, Nuno Paulino, Alex Alcocer, and other
friends not listed here.
I wish you all the best.
Sajjad Fekri
December 20051
1
The research work was carried out at the DSORL laboratory of the ISR, IST, UTL, with
financial support from the Portuguese FCT-POSI programme under framework QCA III, as well as
through PhD grant Ref. SFRH/BD/6199/2001 of the FCT (Portuguese Foundation for Science and
Technology).
xxiv
Chapter 1
Introduction
1
2
methodology for adaptive control of LTI systems, especially for multi-input multi-
output (MIMO) systems, which is based on relevant theoretical results and suitable
for fully addressing the following engineering issues always present in real-life control
systems.
The need for any type of adaptive algorithm arises from uncertainties in the sys-
tem to be controlled. Moreover, the relative importance of estimation and control
problems depends on the degree of uncertainty [4,5]. While uncertainty, in any form,
is the major reason for feedback control, almost all of the classical adaptive control
methods, such as model-reference adaptive control (MRAC), suffer from the fact that
they may lead to instability when there are persistent unmeasurable disturbances,
sensor noise, and significant unmodeled dynamics [6–9]. This means that conven-
tional adaptive algorithms are not robust with respect to dynamic uncertainties in
the system and cannot deal with complex high-dimensional, multivariable systems.
Compensators, which are designed without taking explicitly into account dynamic
and parameter uncertainties, may perform quite poorly in a number of applications.
When using conventional adaptive control, the presence of large parameter errors
will generally result in slow speed, with large transient errors. In a related context,
elegant theoretical results are available but are usually based on very restrictive as-
sumptions (such as minimum phaseness, plant relative degree, positive-realness, and
sign of the high-frequency gain). These results do not yield stability unless these
assumptions are satisfied [8, 10–12]. Thus, adaptive compensators designed based
on these assumptions can be very non-robust, as shown by the well-known counter-
example of [7], demonstrating that the presence of bounded and persistent distur-
bances may lead to the failure of the adaptive scheme through the degradation of
3
performance or even the loss of closed-loop stability. The results of Rohrs et al. [7]
convinced the adaptive control community that existing adaptive control algorithms
could break into instability. They also pointed out that we needed new tools for un-
derstanding the mechanisms of stability and instability in adaptive control systems.
Soon a new field of research on the Robust Adaptive Control Problem was born [9,11].
Research on this topic is vigorously pursued by many distinguished researchers at
present; nobody as yet has arrived on a simple modification to the original adaptive
algorithms that preserves global stability and robustness to unmodeled dynamics,
unmeasured disturbances, and sensor noise.
Another undesirable problem in MRAC strategies is the need to use persistent set-
point excitation. These signals are required to speed-up the identification process and,
hopefully, lead to improved performance (including disturbance-rejection). However,
these excitation signals represent additional disturbances, and they may degrade the
disturbance-rejection of the adaptive feedback systems. Moreover, these strategies
are very sensitive to unmeasured disturbances.
design before implementing the full-blown adaptive design. Also, its limitations must
be explicitly stated (theories have limitations, stupidity does not!)
During the last two decades, there have been great advances in the theory of
robust uncertainty-tolerant multivariable feedback control system design. These non-
adaptive robust algorithms, however, can only meet specifications reflecting low per-
formance in the presence of large parameter uncertainties [1–3, 13, 14]. If all parame-
ters of the system are known, standard control system design methods, e.g. H2 or H∞ ,
are acceptable designs. But if the parameters are uncertain, it is no longer possible
to carry out the designs for lack of a convenient optimization problem formulation.
Multiple-Model Adaptive Estimation (MMAE) and Control (MMAC) methodolo-
gies are model-based adaptive strategies which incorporate a set of model/compensator
pairs rather than relying on a single model and compensator; see, e.g., [5, 15–17].
These methodologies can be used to handle parametric uncertainties and have been
used for many estimation and control applications such as failure detection, adaptive
filtering, large space structures control, target tracking, aircraft control, vibration
control, power systems, and biomedical problems [18–26]. One of the purposes in this
chapter (and also of Chapter 2) is to do a comparative survey of these methodologies.
From the above, we can conclude that there is a pressing need to develop robust-
adaptive compensators in the face of large uncertainties and unmodeled dynamics,
for both SISO and MIMO feedback control systems. Multiple model adaptive control
schemes offer the potential of improved performance over conventional single model
schemes in a number of cases of practical interest. Moreover, there is great potential
for the development of control tools if recent results of non-adaptive robust control
(using µ-synthesis) could be integrated with MMAC methodologies.
Figure 1.1 shows the classical MMAE and the Classical MMAC (CMMAC) archi-
tectures; it is assumed that a linear time-invariant plant is driven by white process
noise ξ(t), as well as a known deterministic input signal u(t), and generates mea-
surements that are corrupted by white noise θ(t); see Chapter 3 for complete details.
ξ (t) θ (t)
ξ (t) θ (t)
u (t ) y (t )
u (t ) y(t )
xˆ1 (t )
r1 (t ) xˆ1 (t ) u1 (t )
r1 (t )
xˆ2 (t ) xˆ(t )
xˆ2 (t ) u2 (t ) u (t )
r2 (t )
r2 (t )
Figure 1.1: The classical MMAE and the Classical MMAC (CMMAC) architecture.
Suppose, for example, that the plant parameter vector in other multiple-model
based control approaches, e.g. switching control scheme illustrated later, changes to
a new vector which is not in the set of models considered. This unanticipated envi-
ronment must be learned by a supervisory logic online. This in turn may make the
convergence slow (in order to achieve small steady-state error, say less than ε), and
the slowness of adaptation may result in a large transient error. This is a major dis-
advantage of switching based control while new plant parameter vectors can be easily
adapted in a probabilistic way using the adaptation philosophy which is inherent to
MMAC algorithms. Instead of using a supervisory structure in MMAC methodology,
a probabilistic weighting evaluator chooses which model best represents the current
plant input/output behavior.
In recent years, adaptive control based on multiple models and model/compensator
switching strategies have attracted considerable attention in a sense that they may be
6
appropriate for realistic applications and more robust to system uncertainties when
compared with classical adaptive schemes, such as MRAC. Such modifications are of
particular importance in many applications where the environment or system param-
eters change abruptly or even if such a system is required to be controlled based on
a single model whose parameters are to be identified from input-output data. This
model will then have to adapt itself to the new environment before appropriate con-
trol action can be applied. In linear systems, such adaptation may be possible, but
the slowness of adaptation may result in a large transient error. In nonlinear systems,
however, a single model may not be adequate to identify the changes in the system,
i.e. a model may not exist in the assumed framework to match the environment.
Traditional adaptive control algorithms focused on parameter identification for
LTI systems subject to known persistent excitations [8, 9]. Convergence proofs, i.e.
parameter identification, of these schemes involved assumptions on relative degree,
sign of high-frequency gain, positive-realness and the absence of unknown (random)
process disturbances and measurement noise. In particular, the classical “model-
reference” adaptive control schemes ignored the presence of inevitable high-frequency
unmodeled dynamics. A counterexample in [7] demonstrated the fragility of these
model-reference schemes in the presence of unmodeled dynamics as evidenced by
long-term instabilities. Ever since the publication of [7] a huge number of ad-hoc
fixes have been proposed to address the “robust adaptive control problem” but none
has found either wide theoretical acceptance or engineering utility. It should also be
noted that almost all such classical adaptive algorithms addressed the single-input
single-output (SISO) case and they are extremely difficult to generalize to the MIMO
problem.
In a completely different vein, the problem of combined estimation and identifica-
tion for uncertain MIMO LTI systems has been addressed since the late 60’s. In the so-
called multiple-model adaptive-estimation (MMAE) method, see [16, ch. 10], [15, 29],
and Figure 1.1(a) – an unknown LTI plant, subject to white noise disturbances and
sensor noise, is postulated to be one of N known models.
Chapter 2 contains a more extensive discussion of recent multiple-model adaptive
methods.
7
probability calculations, and a bank of robust LTI compensators, each designed via
mixed-µ synthesis. Figure 1.2 shows our proposed RMMAC architecture.
ξ (t) θ (t)
Controls Measurements
u (t ) y (t )
Unknown plant
S1 P1 (t )
Posterior
Residual S2 Probability P2 (t )
covariances Evaluator
SN
(PPE) PN (t ) Posterior probabilities
It is emphasized that RMMAC allows the designer to evaluate a-priori the benefits
of using adaptive control as compared to the best non-adaptive MIMO feedback design
with guaranteed stability- and performance robustness. Such a comparison can be
readily accomplished either in the frequency-domain (using maximum singular value
plots) or by comparing RMS errors.
Furthermore, RMMAC feedback designs can be used for the MIMO uncertain
systems with full accounting of model parameter errors, unmodeled dynamics, plant
disturbances, sensor noise and stability- and performance-robustness guarantees. As
we shall elaborate in the sequel, the bank of KFs represents the identification side of
the RMMAC. The proper design of each KF in the RMMAC architecture of Figure 1.2
is crucial in order to satisfy the theoretical assumptions [44, 45, 45] which will imply
that the “Posterior Probability Evaluator” (PPE) will lead to the correct model iden-
tification. Chapter 3 presents the concepts leading to the on-line generation of the
10
posterior probabilities by the PPE and encompasses the key equations for calculating
the Baram Proximity Measure (BPM) to ensure asymptotic convergence of the pos-
terior probabilities. The design of the KFs must be done after the number of models
and their boundaries have been established using the design procedures in Chapter 5,
for a scalar parameter case, and Chapter 8, for the m-dimensional parameter case.
The bank of the “local” mixed-µ compensators in Figure 1.2 are designed using
the state-of-the-art robust mixed-µ synthesis [1, 2, 46, 47]. These local “robust” com-
pensators, referred to as the “local non-adaptive robust compensators (LNARCs)”,
are designed so as to guarantee local stability- and performance-robustness for each
parameter subset (model) covered by the KFs. Each local mixed-µ compensator
generates a local control signal uj (t). The “global” control is then generated, as in
the CMMAC case, by the probabilistic weighting of the local controls uj (t) by the
posterior probabilities Pj (t).
In all the proposed multiple-model adaptive methods, including the RMMAC, the
complexity of the adaptive feedback system directly depends on the number of models
employed, N , and the level of the desired performance improvement. By decreasing
the size of the real parametric subsets to yield high-performance adaptive systems one
obtains more models and hence the complexity of the adaptive system will increase.
Thus, all multiple-model approaches must address the following:
(a) how to divide the initial large parameter uncertain set into N smaller param-
eter subsets,
(b) how to determine the “size” or “boundary” of each parameter subset, and
(c) how large should N be? Presumably the “larger” the N, the “better” the
performance of the adaptive system is.
Note that an RMMAC design has greater complexity than CMMAC, because of
the on-line operation of both the KF dynamics and of the mixed-µ dynamic compen-
sators. In summary, the trade-off between the desired performance and the number
of models/compensator must be done. This is one of the key contributions of this
dissertation.
11
• RMMAC integrates, for the first time, results from stochastic dynamic hypoth-
esis testing (using multiple-model adaptive estimation) and robust compensator
design (using mixed-µ synthesis).
• The RMMAC design methodology can deal with unmeasurable stochastic plant
disturbances, unmeasurable measurement noise, unmodeled dynamics, and one
or more uncertain real parameters (constant or slowly-time varying).
• The precise specification of each of the multiple models in the RMMAC ar-
chitecture is the outcome of explicit performance optimization for the posed
performance specifications.
• The RMMAC methodology is valid for both SISO and MIMO case with full
accounting of model parameter errors.
(b) surveillance, tracking, and other estimation systems that utilize the Extended
Kalman filter (EKF), and
(c) surveillance, tracking systems, and other estimation systems that are based
on the sum-of-gaussian multiple model estimation method, when the extended
Kalman filter is not “good enough”.
None of the above “practical” methods is backed by solid theory. However, all
utilize reasonable extensions of theoretical results valid for linear systems, e.g. the
Kalman filter. The gain-scheduling method also pieces together linear compensators
13
in a common sense manner. The trouble is that most real applications involve non-
linear systems. Clearly, their engineering utility is extensive. Our perspective is that
our proposed highly nonlinear proposed RMMAC is akin to the above. Who knows,
it may become a routine design method in the future.
The structure of this thesis is as follows. In Chapter 2 a brief survey of the history
of multiple model methods is provided including the classical multiple-model adap-
tive estimation (MMAE) that represents the identification part of our proposed robust
adaptive methodology in this thesis. The problem formulation of the MMAE philos-
ophy is discussed in Chapter 3 including some preliminary remarks together with
details on the adaptation of MMAE filter to closed-loop control, so-called MMAC. In
Chapter 4, brief features of the employed mixed real/complex µ-synthesis technique
are presented and an overview of robust control is described.
14
2.1 Introduction
The purpose of this chapter is to present a literature overview that will facilitate the
understanding of the material presented in the main body of the multiple-model based
estimation and control approaches. The basic structure of the chapter is as follows.
In Section 2.2 an overview on classical MRAC is outlined. In Section 2.3 different
multiple-model based adaptive methodologies are discussed including the classical
and switching MMAC methods. Our proposed architecture, associated with robust
multiple-model adaptive control methodology (RMMAC) is introduced in Section 2.4.
Finally, some chapter concluding remarks are discussed in Section 2.5.
15
16
Identification
(Real-Time)
Parameter Adjustment
Logic
Disturbances Sensor
noise
Commands Controls Outputs
Controller Dynamic System
(Compensator) (Plant)
compensator from a conventional one. The identification block contains some kind of
recursive estimation algorithm which aims at determining a best model of the process
at the current instant. The parameter adjustment logic block then uses this model to
produce a compensator according to some design strategy. The control and output
signals of the partially-known dynamic system are monitored, and refinements in the
plant parameters are calculated in real-time. The improved parameter estimates are
then used to modify the characteristics of the dynamic compensator, i.e. simulta-
neous identification and control, so as to improve closed-loop performances, namely
superior command-following and disturbance-rejection.
can become unstable in the presence of plant disturbances, sensor noise and high-
frequency unmodeled dynamics [7]. Moreover the MRAC methodology was limited
to single-input single-output (SISO) plants; attempts to extend the MRAC method-
ology to the multi-input multi-output (MIMO) case were extremely cumbersome.
Because of these shortcomings, we shall not discuss further the MRAC methodology
in the sequel.
Figure 2.2: The large real parameter set is decomposed into N smaller subsets (mod-
els) where the original set is covered by one or more of the models in MMAC based
designs.
One of the first uses of multiple-models was motivated by the need for accurate
stochastic state-estimation for dynamic systems subject to significant parameter un-
certainty. In such problems the estimation accuracy provided by the standard KFs
was not adequate. For some early references regarding MMAE see [5,15,16,29,48–51].
The classical MMAE architecture shown in Figure 2.3 involves the parallel oper-
ation of N discrete-time KFs (each matched to one of the postulated models). By
using a Posterior Probability Evaluator (PPE), the residuals (innovations) of the KFs
can be used to calculate in real-time changes in the posterior probability of which one
of the N models is the correct one. Full details of the MMAE algorithm can be found
19
ξ (t) θ (t)
u (t ) y (t )
xˆ1 (t )
r1 (t )
xˆ2 (t ) xˆ(t )
r2 (t )
xˆN (t)
rN (t)
S1 P1 ( t )
S2 P2 ( t )
SN PN (t )
in Chapter 3 though its notations and formulas are summarized in the sequel.
In the classical MMAE the physical (unknown) plant is assumed to be linear time-
invariant (LTI)1 driven by white process noise , as well as a known deterministic input
signal, and generates measurements that are corrupted by white noise. The plant is
assumed to belong to a basket of N known models, but we do not know at the initial
time t = 0 which one is the correct model. Thus, at time t=0 each one of the N
1
models has the same prior probability Pi (0) = N
of being the true one.
A bank of N parallel KFs is constructed where each KF is matched to the cor-
responding model in the basket of N known models. The objective of developing the
MMAE architecture is to identify an unknown plant and estimate its state variables.
Each KF is driven by the same deterministic control applied to the unknown plant as
well as by the noisy measurements generated by the unknown plant. Then each KF
generates a local state estimate x̂i (t) and a residual signal ri (t), together with state-
covariance and residual-covariance matrices (the covariance matrices are generated
1
The MMAE concept is also available for the linear time-varying case. However, in this thesis
we focus upon the time-invariant version which uses LTI (steady-state) Kalman filters.
20
off-line for this linear case). It is then possible to evaluate in real-time the posterior
probability Pi (t) that each of the N models is indeed the unknown plant, given the
measurements and controls, up to time t. The optimal global state estimate x̂(t) is
obtained by weighting the individual KF state-estimates by the respective posterior
probabilities Pi (t), that each of the N models is indeed the unknown plant, given the
measurements up to time t.
The classical MMAE algorithm is optimal (under the usual linear-Gaussian as-
sumptions) because the conditional probability density function of the state is the
sum of Gaussians, and hence the MMAE generates the true conditional mean and
covariance matrix of the state. It has been shown under these assumptions as well as
the stationarity of the stochastic signals that the correct model is identified “almost
surely” [16, 44, 45, 52]. If the actual system is not identical to one of the postulated
models, then the posterior probability converges to the “nearest probabilistic neigh-
bor” where the distance between two stochastic dynamic systems is based on the
Baram Proximity Measure (BPM); see [44] and Chapter 3.
From a technical point of view, the MMAE system of Figure 2.3 blends opti-
mal estimation concepts (i.e. Kalman filtering) and dynamic hypothesis-testing con-
cepts that lead to a system identification algorithm. As will be explained in Chap-
ter 3, each KF generates a “local” state estimate, x̂k (t|t) and a residual (or inno-
vations) signal, rk (t), which is the difference between the actual measurement and
the predicted measurement (the residual is precisely the prediction error common
to all adaptive systems). Furthermore, the (steady-state) residual covariance ma-
trix, Sk , k = 1, 2, . . . , N , associated with each KF can be computed off-line. The
key to the MMAE algorithm is the so-called “posterior probability evaluator (PPE)”
which calculates, in real time, the posterior conditional probability that each model
generates the data, i.e.
The equations above indicate that the MMAE filter updates the state-estimate and
covariance matrix at any time a new control is applied or a new sensor measurement
is obtained.
The MMAE is valid for both the time-varying as well as the time-invariant (steady-
state) case. In this thesis, we shall concentrate upon the steady-state case using LTI
22
Kalman filters. As shown in Figure 2.3, each steady-state KF generates a local state-
estimate vector and a residual vector in real time. The residual vectors are used to
generate (on-line) the posterior probabilities by the following recursive formula to
indicate which KF model indeed best represents the true uncertain plant.
1
βk e− 2 wk (t+1)
Pk (t + 1) = Pk (t) (2.3)
P
N
− 21 wj (t+1)
Pj (t)βj e
j=1
where
After computing the posterior probabilities Pk (t) for all models at time t, the state
estimate and the state covariance matrix of the MMAE are computed, respectively,
as follows utilizing “probabilistic weighting” by the computed posterior probabilities,
Pk (t) – see eq. (2.3).
N
X
x̂(t|t) = Pk (t)x̂k (t|t) (2.8)
k=1
XN · µ ¶µ ¶T ¸
Σ(t|t) = Pk (t) Σk (t|t) + x̂k (t|t) − x̂(t|t) x̂k (t|t) − x̂(t|t) (2.9)
k=1
The key property of the MMAE algorithm – see Chapter 3 – is that, under suitable
assumptions, one of the posterior probabilities, say Pj , Pj (t) → 1, where j indexes
the model that is “closest” to the correct hypothesis H = Hj , even though the ac-
tual plant parameter is different than pj , as t → ∞. These asymptotic convergence
23
The global feedback control, u(t), is calculated by weighting each local control,
uk (t), by its corresponding posterior probability, Pk (t), i.e.
N
X
u(t) = Pk (t) uk (t) (2.11)
k=1
24
Process Noise
ξ (t) θ (t)
Sensor Noise
u (t ) y (t ) : Measurements
unknown plant
LQ-gains
KFs xˆ1 (t ) u1 (t )
KF #1 r1 (t )
-G1
xˆ2 (t ) u2 (t )
KF #2 r2 (t )
-G2 u (t )
S1 P1 ( t )
Posterior Posterior
Residual S2 Probability P2 ( t )
hypotheses
Covariances Evaluator probabilites
SN
(PPE) PN (t )
Figure 2.4: The Classical Multiple-Model Adaptive Control (CMMAC) Scheme. The
“local” KF state estimates get multiplied by the “local” LQ gains Gk to form the
“local” controls which, in turn, generate a probabilistically weighted “global” control.
Ref. [69] addressed a method to design the representative model set in the MMAE
framework. It also proposed an analogous method for designing MMAC regulator to
provide stabilizing control in the presence of an uncertain real parameter vector by
optimizing a cost functional representing the average regulation error autocorrelation,
with the average taken as the true parameter ranges over the admissible parameter
set. However, this work did not deal with any unmodeled dynamics. It was based
on the numerical Monte Carlo (MC) results and not on any theoretical evidence, as
expected, since the closed-loop MMAC is highly nonlinear even if LQ gains are simply
used. Besides, extending this method to multivariable systems with two (or more)
uncertain parameters seems to be very hard to follow.
25
It was shown that the classical MMAC method does not satisfy an optimality prop-
erty in a stochastic LQG sense [70]. Furthermore, there is no guarantee of robustness
against plant parameter uncertainty for the simple fact that LQG compensators lack
that property as well [71]. Furthermore, any errors in the local state estimates im-
mediately influence the controls. Thus, there is no “separation property” between
“system identification” and “control” in CMMAC architectures. This can be viewed
as one of the main disadvantages of the CMMAC.
The CMMAC methodologies did not reflect later advances in the robust control of
linear dynamical systems. In the context of this chapter, it is important to stress that
no robustness to unmodeled dynamics was considered in early CMMAC designs (such
robustness issues were unknown in the 1970’s) and that the performance specifications
were generated by LQG “tricks” and not with the frequency-weight concepts (H2 and
H∞ ) widely adopted at present. Thus, there was a great potential for the development
of control tools if the recent results of non-adaptive robust control (using the mixed-µ
synthesis) could be successfully integrated in MMAC methodologies. It remained an
open research question on what is the proper way to integrate recent robust control
results with the classical MMAC algorithms, which was in fact the motivating purpose
of this thesis.
It is therefore relevant to combine the benefits of MMAC strategies with those
of non-adaptive robust µ-synthesis control, by combining the theoretical approaches
of multiple-model robust-adaptive control methodologies with mixed-µ synthesis in
the presence of unmodeled dynamics and large uncertainties in the system. This
is discussed in Section 2.4. It should also be noted that the closed-loop MMAC
system is highly nonlinear, and hence very hard to analyze from a formal point of
view [18, 65–67].
The use of multiple models to switch or reset parameter estimators has been proposed
in order to speed up the convergence rate in the context of certainty equivalence
adaptive control of linear systems.
26
Figure 2.5 shows a block diagram of the switching control scheme, which in fact
is a logic-based switching control scheme, where the controlled variable y and the
control variable u are measured and used for the real-time evaluation.
ξ (t)
.
.
The recent literature in switching based multiple-model adaptive control has con-
sidered approaches such as adaptation using multiple models (see e.g. [33–35]), unfal-
sified control (see e.g. [40–43,72]), and switching based multiple model adaptive control
(see e.g. [31, 32, 37–39, 73, 74]). A brief overview of these switching control schemes is
presented in the sequel. However, detailed analyses and discussions are beyond the
scope of this thesis.
In the so-called adaptation using multiple models method in order to control linear
(and possibly nonlinear) systems, usually SISO, a finite number of models/compensators
are used. The (unknown) parameter vector at the models are initialized as some pre-
defined vectors and updated during the control process.
Ref. [33] has studied an adaptive control scheme using multiple models in order
to improve the command-following transient response. The idea is similar to the
standard model reference adaptive control (MRAC) problem. The proposed adaptive
control system consists of N compensators and N identifiers where each identifier
corresponds to a compensator. The structure of the identifiers is the same but with
27
different initial estimates of the plant parameters. For example, a sample switching
scheme among the compensators (in order to choose the best compensator to achieve
the better performance) adopts the quadratic criterion function [75]
Z t
Jj (t) = αe2Ij (t) +β e2Ij (τ )dτ, (α, β > 0) (2.12)
0
where the asterisk (∗) denotes the convolution operator, r̃j (t) is the fictitious input
associated with compensator #j, and ω1 (τ ) and ω2 (τ ) are the impulse responses of
filters ω1 (s) and ω2 (s).
Figure 2.6 shows the unfalsified control architecture including the representation
of actual and equivalent closed-loop systems.
If the current compensator is judged to be unable to provide satisfactory per-
formance, it is falsified (i.e., discarded) and replaced by the candidate compensator
that minimizes the performance index. Thus, switching control is closely related to
unfalsified control. It infers the behavior of a feedback loop consisting of an uncer-
tain plant and a given compensator using data obtained from the same plant and
possibly a different compensator. Both unfalsified control and switching control have
29
(a) Actual single closed-loop system (b) Equivalent single closed-loop system
Unfalsified signal
~
r1(t)
K1
~
r2(t)
K2 Unfalsification Algorithm
for candidate controller
(u,y)
. . .
~
ri(t)
Ki
u(t) y(t)
.
Unknown Plant
. . .
.
.
r~N(t)
KN
Controllers
In recent years, there has been considerable research activity in switching multiple
model adaptive control systems for solving adaptive control problems. A switching
control scheme is typically composed of an inner loop where a candidate compensator
is connected in closed-loop with the system, and an outer loop where a supervisor,
30
based on input-output data, decides which compensator to place in feedback with the
system and when to switch to a different one.
Logic based switching compensators were first proposed in [28] and later turned
out to be a valid alternative to the more traditional continuously tuned adaptive com-
pensators. The compensator switching strategy is based on the real-time switching
among a set of candidate compensators. This candidate set is selected a priori based
on the control objectives and available process knowledge such as a nominal com-
pensator and the range of process operating conditions that the control system must
accommodate. The evaluations can be performed at every sampling instant, less fre-
quently, or upon request. In general, when the uncertainty on the system description
is large, no single candidate compensator in a given family is able to adequately reg-
ulate all the admissible models for the system. This motivates the use of a switching
control scheme where a supervisor decides on-line, based on the observations collected
from the operating system, which is the best compensator to be applied and, when if
necessary, to switch to a different compensator.
Here, the SMMAC architecture is briefly discussed to point out some similar-
ities and differences to the CMMAC of Figure 2.4 and the RMMAC architecture
to be discussed below. The approach is deterministic and the goal is to prove “lo-
cal” bounded-input bounded-output stability of the SMMAC system under certain
assumptions. The plant-disturbance and sensor-noise signals are assumed bounded
rather than being characterized as stochastic processes as in CMMAC and RMMAC.
The presence of unmodeled dynamics is also considered, although the bound on the
31
ξ (t) θ (t)
u (t ) y (t )
e1 (t ) u1 (t )
e2 (t ) u2 (t ) u (t )
eN (t ) u N (t )
µ1 (t )
µ2 (t ) σ k (µ k )
µN (t)
unmodeled dynamics is simply an H∞ bound. The theoretical results to date are re-
stricted to single-input single-output (SISO) systems, although research is underway
to extend them to the MIMO case.
In reference to Figure 2.7, the SMMAC employs a finite number of stable deter-
ministic estimators (Luenberger observers), called multi-estimators and denoted by
Ek (s); k = 1, 2, . . . , N , designed for a grid of distinct parameter values pk ∈ P , where
P is the compact real-parameter set. The output of each estimator yk (t) is compared
with the measured (true plant) output y(t) to form the estimation (prediction) errors
ek (t) = yk (t)−y(t); k = 1, 2, . . . , N . We remark that the errors ek (t) in Figure 2.7 are
completely analogous to the residuals rk (t) in the MMAE (Figure 2.3), the CMMAC
(Figure 2.4) and the RMMAC (Figure 1.2). The monitoring signal generator, M (s),
is a dynamical system that generates monitoring signals µk (t); these are suitably-
defined integral norms of the estimation errors ek (t). The size of these monitoring
signals indicates which of the multi-estimators is “closer” to the true plant. In addi-
tion to the bank of multi-estimators, it is assumed that a family of multi-compensators,
Ck (s); k = 1, 2, . . . , N , has been designed so that each provides satisfactory stable
32
feedback performance for at least one discrete parameter pk ∈ PD . The basic idea
is to use the monitoring signals µk (t) to “switch-in” the suitable compensator. This
is accomplished by a switching logic S which generates a signal σ(t) ∈ PD that can
be used to switch-in the appropriate compensator. A key property of the switching
logic S is that it keeps its output σ(t) constant over some suitably long “dwell time”;
this avoids rapid switching of the compensators and allows most transients to die-out
between compensator switchings. The details of the switching logic differ in the cited
SMMAC references.
The basic structural difference between the CMMAC and the SMMAC – compare
Figures 2.4 and 2.7 – is that in the SMMAC the “identification” process is com-
pletely separated from the “control” process. Even in the case of the CMMAC that
the largest posterior probability switches the corresponding LQ control gain, the iden-
tification and control get “mixed-up”. The separation of the identification and control
processes in the SMMAC seems to have an advantage, coupled with the idea of in-
frequent compensator-switching. Otherwise, the KFs in the CMMAC serve the same
objective as the multi-estimators of the SMMAC. Moreover, the multi-compensators
in the SMMAC can be more complex than the simple LQ-gains in the CMMAC.
Unfortunately, SMMAC numerical simulations have been reported for only a couple
of (very) academic SISO plants.
Refs. [28, 81] introduce a simple switching logic called “hysteresis switching”. In
this hysteresis switching logic, a switching occurs when a monitoring signal that
corresponds to the compensator placed currently in the feedback loop exceeds the
smallest monitoring signal by a prespecified positive number, called the hysteresis
constant. This logic, under some assumptions, guarantees that for a finite family of
monitoring signals switching stops in finite time.
Refs. [32, 78] have studied the use of multiple fixed models for an SISO system.
The author has introduced the dynamics of a multi-estimator, a performance-weight
generator and a dwell-time switching logic. By comparing the prediction errors of
the fixed models and switching based on the “certainty equivalence principle”, the
author has defined a supervisory compensator, and showed the tracking performance
and robustness of the control system. However, the multiple fixed models need to
33
be chosen with great care. Besides, in spite of the author’s claim that the designed
supervisor performs remarkably well in simulation even if the uncertain process model
parameter changes slowly with time, our simulations of the method showed long-term
instabilities.
Ref. [37] has outlined the problem of constructing an appropriate model set for
which to design a family of linear candidate compensators that serve as candidates
for a multiple model switching adaptive control scheme. In this work, the authors
argue that the Vinnicombe metric is more control relevant than the gap metric. This
led them to use a generalized closed-loop sensitivity transfer function matrix to de-
sign the finite dimensional compensators set based on this metric. The Vinnicombe
distance (or ν-gap metric) calculated for all model/compensator pairs shows that
which pair is the best candidate. Ref. [38] has discussed the problem of control of
an uncertain continuous-time SISO linear system using switching logic. They have
discussed in design and analysis of hysteresis-based switching control algorithms. The
switching-logic based control methodology is relied on “estimator-based” and similar
to the obtained results in [32]. Ref. [39] has tried to bring together the two con-
cepts of multiple model adaptive control and “safe adaptive control” and proposed
an algorithm in which a closed-loop transfer function must be accurately identified.
The work has outlined a multi-model adaptive control algorithm to achieve the safe
adaptive control properly. That is, if the compensator, introduced at any time, is
unchanged from this time on, the resulting time-invariant system (closed-loop) would
be stable. This method involves a frequency-dependent performance measure and
also employs the Vinnicombe metric.
Because of conventional adaptive control shortcomings, it has been very difficult
to develop commercial adaptive control systems that are robust, perform well, and
can be effectively used by typical plant personnel. Consequently, considerable effort is
required to ensure that commercial compensators will be suitable for industrial appli-
cations. Ref. [27] has presented some procedures in order to improve the robustness of
adaptive compensators for those applications. It is suggested to use supervision based
adaptive compensator in industrial control systems where the compensator settings,
or even the control configuration, are adjusted on-line as process conditions change.
34
Some related work has discussed convergence and stability of the switching control
scheme for adaptively stabilizing an unknown system for some classes of systems,
most of them linear and affected by noise; see e.g. [82–86].
ξ (t) θ (t)
u (t ) y (t )
r1 (t ) u1 (t )
K1 ( s)
r2 (t ) u2 (t ) u (t )
K 2 ( s)
rN (t ) u N (t )
K N (s )
S1 P1 (t )
S2 P2 (t )
SN PN (t )
Figure 2.8: The architecture of the proposed RMMAC feedback system. The Kalman
filters (KFs) generate the residuals rk (t). The PPE uses both these residuals and the
constant residual covariances Sk to generate the posterior probabilities Pk (t) that
weigh the local robust controls uk (t) generated by the LNARCs which are designed
using the mixed-µ methodology.
By comparing Figures 2.8 and 2.3, the basic idea behind the proposed RMMAC
35
architecture can be seen. First, the bank of KFs is used to generate via its residuals
a posterior probability of which model is most likely (but not for state estimation
as in the classical MMAE and CMMAC methods). Second, a probabilistic weighting
of each “local” robust control signal is used to implement the global control applied
to the plant. Also, by comparing Figures 2.8 and 2.4, it is clear that RMMAC is
very different from classical MMAC, because in a RMMAC system the local KF
state-estimates are not explicitly used ; the MMAE only acts as a system identifier.
The physical plant is assumed to belong to a “legal family” of possible plants,
where the nominal plant together with frequency-dependent upper bounds on un-
modeled dynamics and upper- and lower-bounds on key uncertain real parameters
defines this “legal family” of plants. The performance specifications are explicitly
stated in the frequency domain; they typically require superior disturbance-rejection
in the lower frequency region while safeguarding for excessive control action at higher
frequencies. At present, robust synthesis can not deal with slowly-varying parame-
ters, from a theoretical perspective. However, as we suggest in Chapter 10, advances
in LPV analyses and designs may be relevant.
It should be self-evident that the parameter uncertainty for each of the “local”
N models is much less than that of the true plant. Therefore, one can design, using
mixed-µ synthesis, “local” robust compensators (LNARCs) for each model and each
will have a better guaranteed performance as compared to the “global” robust de-
sign for the much more uncertain true plant. This led us to construct the proposed
RMMAC architecture as shown in Figure 2.8.
This dissertation focuses on linear time-invariant systems affected by parametric
uncertainties where the multiple-model adaptive estimators are used to handle para-
metric uncertainties. We introduce a robust multiple-model adaptive control (RM-
MAC) methodology based on µ-synthesis. The operating range of a large uncertain
parameter is decomposed into a number of small uncertain areas; see Figure 2.9.
These can then be used to represent the designed Robust MMAC system using as-
sociated simple subsystems (called models in the MMAC methodology). Each model
has also unmodeled dynamics. Much of the thesis is concerned with how to deter-
mine these models, how to incorporate non-adaptive robust control strategy using
36
Figure 2.9: RMMAC philosophy in the two-dimensional parameter case. The large
bound in the plant real uncertainties is decomposed into N small model spaces.
µ-synthesis into the classical MMAC methodology for each of the models, and how to
integrate the µ-synthesis stability-performance issues together with the adaptation of
the MMAC methodology.
It is very desirable that the above attributes of nonadaptive robust feedback sys-
tems be also reflected in the design of robust adaptive compensators as well. Thus, in
our view, an adaptive control design must explicitly yield stability- and performance-
robustness guarantees, not just stability (which has been the central focus of almost
all adaptive control methodologies).
The key characteristics of the RMMAC, as discussed in Chapter 1, are again
summarized below.
• RMMAC integrates, for the first time, results from stochastic dynamic hypoth-
esis testing (using multiple-model adaptive estimation) and robust compensator
design (using mixed-µ synthesis).
• The RMMAC design methodology can deal with unmeasurable stochastic plant
disturbances, unmeasurable measurement noise, unmodeled dynamics, and one
or more uncertain real parameters (constant or slowly-time varying).
37
• The precise specification of each of the multiple models in the RMMAC ar-
chitecture is the outcome of explicit performance optimization for the posed
performance specifications.
• The RMMAC methodology is valid for both SISO and MIMO case with full
accounting of model parameter errors.
To the best of our knowledge, none of the adaptive methods (including CMMAC,
SMMAC, and RMMAC) have been proven to be globally stable under a variety of
operating conditions. The basic difficulty is that all of these methods lead to a highly-
nonlinear time-varying (stochastic) closed-loop system, even in the case of constant
but unknown real parameter(s). The difficulty increases when the parameters vary
with time. There does not exist a sufficiently general stability theorem that can be
adapted to prove global adaptive control stability. The available results relate to
38
(b) surveillance, tracking and other estimation systems that utilize the extended
Kalman filter, and
(c) surveillance, tracking systems and other estimation systems that are based
on the sum-of-gaussian multiple model estimation method, when the extended
Kalman filter is not good enough.
These most useful “practical” design methods are not based upon solid global
theory. However, all exploit reasonable extensions of theoretical results valid for linear
systems. The gain-scheduling method [91–93] also pieces linear designs in a common
sense manner2 . Our perspective is that the highly nonlinear proposed RMMAC is
akin to the above and it is in this spirit that we offer the RMMAC methodology to
the literature.
them in a fair manner it is suggested that all multi-compensators of the SMMAC de-
signs should utilize the available results and software associated with mixed-µ synthe-
sis. If the specifications for required stability-robustness and performance-robustness
for the adaptive closed-loop system are explicitly stated, then one can evolve the
newer RMMAC architectures so that they can be fairly compared in several realistic
applications.
This chapter concludes with a few comments. Even though all MMAC architec-
tures are made by piecing together LTI systems, the probabilistic weighting in the
RMMAC, as well as the switching logic of the SMMAC, result in a highly nonlinear
and time-varying closed-loop MIMO feedback system. Even in the simpler CMMAC,
involving LQG compensators, attempts to prove global stability were not success-
ful [65–67, 70]. Hence, it is wishful thinking to expect foolproof global asymptotic
stability results in the near future, because there does not (as yet) exists a solid
mathematical theory for global stochastic nonlinear time-varying asymptotic stabil-
ity.
40
Chapter 3
3.1 Introduction
The purpose of this chapter is to overview the basic results from probability and
estimation theory that will be used in the thesis. First, a brief review on the classical
Multiple-Model Adaptive Estimation (MMAE) is given that is restricted to linear
systems driven by white noise Gaussian inputs. The classical MMAE assumes that
the true model is included in the model set, as MMAE can identify it under a verifiable
uniqueness condition. If the true model is not included in the model set, then the
identification procedures converge to a model in the set which is “closest” to the true
model in an information metric sense, as proved by Yoram Baram (see [44]); we shall
also summarize his results in this chapter. In that case, the posterior probabilities of
models will converge if certain ergodicity and stationarity assumptions of residuals
(innovation signals) hold. These conditions are also summarized in this chapter.
The structure of the Chapter is as follows. Section 3.2 is a brief summary to
the discrete-time Kalman filter. Section 3.3 discusses notations and some important
characteristics of the classical MMAE filter when the unknown plant belongs to a
model set. Section 3.4 presents the convergence of the maximum likelihood estimate in
the presence of modeling error when the classical MMAE does not apply. Section 3.5
provides a description of the Baram Proximity Measure (BPM) [44]. Special attention
41
42
is paid to the role of the probabilistic ergodicity assumptions and to the formulation
of the identification problem. The basic result is a convergence result for model
identification. It should be noted that similar results have been proved in [16] using
the Kullback information distance. The results of this chapter will be used in the
RMMAC design methodology in Chapters 5-9. Section 3.6 contains the remarks and
conclusions.
where x(t) is the state, u(t) is the deterministic input, t represents the discrete
time variable, the plant white noise is ξ(t) with zero mean and intensity Q and the
sensor white noise is θ(t) with zero mean and intensity R. The two noise sources are
assumed independent of each other and independent of the input.
The task of the Kalman filter can now be stated as: Given a system such as the
one shown in eq. (3.1), how can we filter y(t) so as to estimate the variable x(t) while
minimizing the effects of ξ(t) and θ(t)?
We use this a priori estimate to predict an estimate for the output, ŷ(t|t − 1).
The notation x̂(t|t − 1) can be read as “an estimate of x at time t, based on the
information from time t − 1”; in other words, the estimate based only upon the past
43
values of the outputs and controls, or the a priori estimate. The notation x̂(t|t) can
be read as “an estimate of x at time t, based on the information up to and including
time t”; in other words, the a posteriori estimate based upon the past and current
values of the outputs and controls.
The difference between the estimated output ŷ(t|t − 1) and the actual output y(t)
is called the residual r(t) (or innovation). If the residual is small, it generally means
we have a good estimate; if it is large the estimate is not so good. We can use the
residual to refine our estimate of x̂(t|t − 1); we call this new estimate the a posteriori
estimate, x̂(t|t). If the residual is small, so is the correction to the estimate. As the
residual grows, so does the correction.
The task is to find the Kalman filter gain matrix K that is used to refine our state
estimate, and it is this process that is at the heart of Kalman filtering. Since we are
trying to find an optimal estimator, so we must use an optimal value for the gain, K.
If the measurement noise intensity, R, is very large, K is very small, so we es-
sentially disregard the current measurement in forming the new estimate. This is
as expected; if the measurement noise is large, then we have low confidence in the
measurement and our estimate will depend more upon the prior estimate.
The pertinent discrete-time steady-state Kalman filter equations for the above
plant model are summarized by the following two cycles:
Predict-cycle:
Update-cycle:
The residual covariance matrix S, and the KF gain-matrix K are constant, since
we only deal with the steady-state version of the Kalman filter.
Residual:
KF gain:
K = ΣC T R−1
Σp = AΣAT + LQLT
Σ = Σp − Σp C T S −1 CΣp
It seems reasonable to achieve an estimate of the state (and the output) by simply
employing an optimal estimation architecture, i.e. the Kalman filter as shown in
Figure 3.1.
x(0) ξ (t)
θ (t)
u (t ) y (t + 1)
r (t + 1) xˆ(t+1|t+1) xˆ(t|t)
K
yˆ(t+1|t) xˆ(t+1|t)
C
B A
to the time-invariant one. It is also remarked that all the notations and the proofs of
this section are “borrowed” from [15].
First, let us assume that N discrete-time plant models are given. The dynamics of
the k-th KF model, including the deterministic input, is given below where the time
index is t = 0, 1, 2, . . . and the model index is k = 1, 2, . . . , N .
and the true conditional covariance matrix of the present state vector, x(t), i.e.
£ ¤
Σ(t|t) = E (x(t) − x̂(t|t))(x(t) − x̂(t|t))0 |Y (t) .
The MMAE filter is driven by the sequence of past controls and noisy sensor
measurements while generates both a state-estimate vector and a corresponding error-
covariance matrix. It is a recursive algorithm and updates the state-estimate and
covariance every time a new control is applied and a new sensor measurement is
obtained.
46
Both the MMAE state-estimate and state-covariance matrix represent true condi-
tional expectation and conditional covariance. This fact will be proven in the sequel.
To prove these assertions, we must explicitly calculate the conditional probability
density function p (x(t)|Y (t)). We shall show that the desired probability density
function (PDF), p (x(t)|Y (t)), turns out to be a weighted sum of Gaussian densities,
where the weights are found from the posterior probability evaluator.
Proof. Hypotheses
Let suppose H indicates the hypothesis random variable (scalar) which can attain
only one of N possible values,
H ∈ {H1 , H2 , . . . , HN } (3.5)
The event H = Hk means that the k-th system is the true one, i.e. the one that is gen-
erating the data inside the “black box”. The prior probabilities, Pk (0) ≡ Prob(H =
Hk ); k = 1, 2, . . . , N , at the initial time t = 0 are known.
N
X
Pk (0) ≥ 0, Pk (0) = 1 (3.6)
k=1
Prob(x = Xk ) = Pk ; k = 1, 2, . . . , N
N
X
p(x) = Pk δ(x − Xk ) (3.8)
k=1
where δ(x − Xk ) is the unit impulse at x = Xk . Note that the area under the the
area of each unit impulse is unity, i.e.
Z Z X
N N
X Z
p(x)dx = Pk δ(x − Xk )dx = Pk δ(x − Xk )dx = 1
k=1 k=1 | {z }
1
The key quality of interest in the estimation problem is the conditional density
function, p (x(t)|Y (t)). Consider the joint density function, p (x(t), H|Y (t)). Using
the marginal density function we have
Z
p (x(t)|Y (t)) = p (x(t), H|Y (t)) dH (3.9)
Substitute eq. (3.10) into eq. (3.9) and use eq. (3.7) to obtain
Z
p (x(t)|Y (t)) = p (x(t)|H, Y (t)) p (H|Y (t)) dH
N
X Z
= Pk (t) p (x(t)|H, Y (t)) δ(H − Hk )dH
k=1
XN
= Pk (t)p (x(t)|Hk , Y (t)) (3.11)
k=1
48
Z
p (x(t)|Y (t)) = p (x(t)|H, Y (t)) p (H|Y (t)) dH
Z N
X
= p (x(t)|H, Y (t)) Pk (t)δ(H − Hk )dH
k=1
N
X Z
= Pk (t) p (x(t)|H, Y (t)) δ(H − Hk )dH
k=1
XN
= Pk (t)p (x(t)|Hk , Y (t)) (3.12)
k=1
But the conditional density p (x(t)|Hk , Y (t)) is precisely the conditional density
of the k-th KF which assumes that H = Hk , i.e. that the true system is the k-th
model. Thus, we know that
Therefore, we see that the desired conditional density p (x(t)|Y (t)) is a proba-
bilistically weighted sum of N Gaussian densities, p (x(t)|Hk , Y (t)), each of which is
generated by the bank of the N KFs. All that remains is to calculate the posterior
probabilities
(x(t) − x̂k (t|t) + x̂k (t|t) − x̂(t|t)) (x(t) − x̂k (t|t) + x̂k (t|t) − x̂(t|t))0 =
(x(t) − x̂k (t|t)) (x(t) − x̂k (t|t))0 + (x̂k (t|t) − x̂(t|t)) (x̂k (t|t) − x̂(t|t))0 +
(x(t) − x̂k (t|t)) (x̂k (t|t) − x̂(t|t))0 + (x̂k (t|t) − x̂(t|t)) (x(t) − x̂k (t|t))0 =
We will have
Z
(x(t) − x̂(t|t)) (x(t) − x̂(t|t))0 p (x(t)|Hk , Y (t)) dx(t) =
Z
(x(t) − x̂k (t|t)) (x(t) − x̂k (t|t))0 p (x(t)|Hk , Y (t)) dx(t) +
| P
{z }
k (t|t)
Z
0
(x̂k (t|t) − x̂(t|t)) (x̂k (t|t) − x̂(t|t)) p (x(t)|Hk , Y (t)) dx(t)+
| {z }
1
50
Z
(x(t) − x̂k (t|t)) p (x(t)|Hk , Y (t)) dx(t) (x̂k (t|t) − x̂(t|t))0 (3.19)
| {z }
0
Z
+ (x̂ − k(t|t) − x̂(t|t)) (x(t) − x̂k (t|t))0 p (x(t)|Hk , Y (t)) dx(t)
| {z }
0
The global covariance matrix Σ(t|t) cannot be precomputed off-line, even though
the (local) KF covariance Σk (t|t) are computed off-line. The posterior probabilities,
Pk (t) = Prob (H = Hk |Y (t)), however, must be computed on-line. The mean correc-
tion terms, (x̂k (t|t) − x̂(t|t)), must also be computed on-line. All that remains is to
derive the recursive relation that generates the posterior probabilities,
Also recall that since H is a discrete random variable, the associated probability
density function is a weighted sum of impulses, i.e.
N
X
p (H|Y (t)) = Pk (t)δ(H − Hk ) (3.21)
k=1
XN
p (H|Y (t + 1)) = Pk (t + 1)δ(H − Hk ) (3.22)
k=1
51
Probability Relations
Probability Calculations, I
In eq. (3.25) we need to evaluate p (y(t + 1)|u(t), Hk , Y (t)). But, for the k-th
model, we can equate
p(y(t + 1)|u(t), Hk , Y (t) is Gaussian with mean (3.27) and covariance (3.28), i.e.,
1 1 0 −1
p (y(t + 1)|u(t), Hk , Y (t)) = p e− 2 rk (t+1)Sk (t+1)rk (t+1) (3.29)
(2π)m/2 det Sk (t + 1)
Probability Calculations, II
In eq. (3.25) we also need to evaluate p (y(t + 1)|u(t), Y (t)). Using the marginal
density and Bayes rule we deduce that
Z
p (y(t + 1)|u(t), Y (t)) = p (y(t + 1), H|u(t), Y (t)) dH
Z
= p (y(t + 1)|H, u(t), Y (t)) p (H|u(t), Y (t)) dH
Z
= p (y(t + 1)|H, u(t), Y (t)) p (H|Y (t)) dH
Z N
X
= p (y(t + 1)|H, u(t), Y (t)) Pj (t)δ(H − Hj )dH
j=1
N
X
= Pj (t)p (y(t + 1)|Hj , u(t), Y (t)) (3.30)
j=1
Then, from eqs. (3.31), (3.32), (3.33), and (3.29), we deduce that the posterior prob-
abilities can be computed on-line by the Posterior Probability Evaluator (PPE) by
the recursive formula
βk (t + 1)e−(1/2)wk (t+1)
Pk (t + 1) = · Pk (t) (3.34)
P
N
Pj (t)βj (t + 1)e−(1/2)wj (t+1)
j=1
53
ξ (t) θ (t)
u (t ) y (t )
xˆ1 (t )
r1 (t )
xˆ2 (t ) xˆ(t )
r2 (t )
xˆN (t)
rN (t)
S1 P1 ( t )
S2 P2 ( t )
SN PN (t )
It is remarked that the MMAE algorithm was initially proposed in the late 60’s
and called “partitioned estimation”, especially in the research by Lainiotis [29].
It can also be seen that the structure of the MMAE algorithm is very simi-
lar to that used in the so-called Sum of Gaussian Methods for nonlinear estima-
tion [16, ch. 8]. Clearly, as in the single-model case, the past control sequence
{u(0), u(1), . . . , u(t − 1)} influences the conditional state-estimate and residuals at
time t. Unlike the single-model case, in the MMAE algorithm, the past control se-
quence {u(0), u(1), . . . , u(t − 1)} also influences the conditional covariance matrix,
Σ(t|t), and hence the accuracy of the state-estimate. This implies that some con-
trol sequences are better for improving the accuracy of the state-estimates and model
identification probabilities. If, say, H = Hi is true, the i-th model is the true one,
then
lim Pi (t) = 1, lim Pj (t) = 0, ∀j 6= i
t→∞ t→∞
point of view. This algorithm is optimum in a probabilistic sense in state and param-
eter estimation if the discretized parameter space indeed contains the true parameter.
This is true because: (1) we use the conditional mean as the estimate and (2) the
algorithm was derived without using any approximations.
3.3.4 Discussion
We now discuss the asymptotic properties of this algorithm from a heuristic point of
view. If the system is subject to some sort of persistent excitation, then one would
expect that the residuals of the KF associated with the correct model, say the i-th one
will be ”small”, while the residuals of the mismatched filters (j 6= i ; j = 1, 2, . . . , N )
will be ”large”.
Thus, if i indexes the correct model we would expect
If such a condition persists over several measurements, eq. (3.34) shows that the
“correct” probability Pi (t) will increase while the “mismatched model” probabilities
will decrease. To see this one can rewrite eq. (3.34) as follows,
" #
1 P 1
Pi (t) (1 − Pi (t)) βi e− 2 wi (t+1) − Pj (t)βj e− 2 wj (t+1)
j6=i
Pi (t + 1) − Pi (t) = (3.36)
P
N
− 12 wj (t+1)
Pj (t)βj e
j=1
On the other hand, for the incorrect model, indexed by j 6= i, the same assump-
tions yield
−Pj (t) Pi (t) βi
Pj (t + 1) − Pj (t) ' <0 (3.38)
P
N
− 21 wj (t+1)
Pj (t)βj e
j=1
so that Pj (t) → 0, i.e. the probabilities associated with the incorrect models will
decrease to approach 0.
The same conclusions hold if we rewrite eq. (3.34) in the form
" #
P ³ 1 1
´
Pi (t) Pj (t) βi e− 2 wj (t+1) − βj e− 2 wj (t+1)
j6=i
Pi (t + 1) − Pi (t) ' (3.39)
P
N
− 12 wj (t+1)
Pj (t)βj e
j=1
The above discussion points out that this ”identification” scheme is crucially de-
pendent upon the regularity of the residual behavior between the ”matched” and
”mis-matched” KFs.
The issue of “identifiability” may also be considered from the perspective of per-
sistent excitation signals that need to be adequate for the system identification. See
e.g. [96] for more discussion. In the MMAE algorithm the plant white noise provides
adequate persistant excitation because in KF theory we assume that the system is
controllable from the plant noise, i.e. [Ak , Lk ] is controllable and [Ak , Ck ] is observable
for all k = 1, 2, . . . , N .
modeling error cases. The previous analysis, the classical MMAE, does not apply to
this situation and the question of convergence of the maximum likelihood estimate
has to be analyzed in the present context. One possible approach to this problem,
suggested by Baram in Refs. [44, 97], relies on an information theoretic concepts
so-called Baram Proximity Measure (BPM). Formal definitions and notations of the
BPM are presented in Section 3.5. Refs. [44,45,52,98,99] contain all detailed notations
and descriptions
In this section, we focus on the maximum likelihood identification method for
determining the parameters of a linear-Gaussian state-space model of a dynamic
system in which the modeling uncertainty is denoted by α. This class of models was
described in Section 3.3. For simplicity, we use the notation of yt+1 instead of y(t+1).
The key idea in the extension of the maximum likelihood method to the identifi-
cation problem is to write the more general factorization
and to recall that (in the linear-Gaussian case) p(yτ |yτ −1 ; α) is characterized com-
pletely by quantities computed by the KF corresponding to α.
It is again emphasized that the residuals (innovations) are the differences between
what comes out of the sensors and what was expected, based on the estimated plant
state. If the plant was perfectly modeled in the KF, the innovations would be a
zero-mean white-noise process and its autocorrelation function would be zero. The
departure of the empirical autocorrelation of innovations from this model is a useful
tool for analysis of mismodeling in real-world applications [95].
The problem is to evaluate the maximum likelihood estimate α̂t of α which, by def-
inition, maximizes p(yt ; α). As mentioned above, this probability density function
factors as follows:
But recall that the conditional probability density p(yτ |yτ −1 ; α) of the current
observations yτ given past observations yτ −1 is (in the case of linear-Gaussian state-
space models):
0 −1 (τ ;α)r(τ ;α)
p(yτ |yτ −1 ; α) = (2π)−m/2 (det[S(τ ; α)])−1/2 e−1/2r (τ ;α)S (3.40)
Since the residuals r(τ ; α) = y(τ )− ŷ(τ |τ −1; α), where ŷ(τ |τ −1; α) and the resid-
ual covariances S(τ ; α) are generated by the KF corresponding to α, the likelihood
function is readily computable for every α and set of data yt .
Now to simplify the manipulation of the above quantities, it is customary to
equivalently maximize the logarithm, ln p(yt ; α) instead of p(yt ; α) itself. This has
the advantage of transforming the products into sums, which will be useful when one
needs to compute derivatives. It also replaces the exponential term in eq. (3.40) by a
more amenable quadratic term. Indeed,
t
X
lnp(yt ; α) = lnp(yτ |yτ −1 ; α)
τ =0
where
p(y0 |y−1 ; α) ≡ p(y0 ; α)
and
m 1 1
lnp(yτ |yτ −1 ; α) = − ln(2π) − ln(det[S(τ ; α)]) − r0 (τ ; α)S −1 (τ ; α)r(τ ; α) (3.41)
2 2 2
Furthermore, note that − m2 ln(2π) is a constant term independent of α. Therefore
the maximum likelihood estimate α̂t can be more simply evaluated by minimizing the
“interesting part” of the negative log likelihood function:
t
X
(t+1)m
ζ(yt ; α) ≡ −[lnp(yt ; α) + ln(2π) 2 ]= ζ(yτ |yτ −1 ; α) (3.42)
τ =0
where
1 1
ζ(yτ |yτ −1 ; α) ≡ ln(det[S(τ ; α)]) + r0 (τ ; α)S −1 (τ ; α)r(τ ; α) (3.43)
2 2
Note that the new log likelihood function ζ(yt ; α) has two parts: a deterministic
part which depends only on S(τ ; α), τ = 0, 1, . . . , t and which is therefore precom-
putable and a quadratic in the residuals part, which therefore depends on the data.
58
The above equations are valid for the general case of a time varying system. Note
that each evaluation of the likelihood function requires the processing of the obser-
vations by a time-varying Kalman filter. A very common practice in the case of time
invariant systems is to use instead the steady-state Kalman filter. This introduces
an approximation into the computation of the likelihood function, but this approxi-
mation will be good if the optimal time-varying Kalman filter reaches steady-state in
a time that is short relative to the time interval of the observations. Of course, use
of the steady-state Kalman filter greatly simplifies the calculation of the likelihood
function.
In [22] two distinct MMAE strategies using either constant-gain or time-varying-
gain KFs are discussed to identify the closest model of a mass-spring system. The
results of the constant-gain and the time-varying-gain MMAE algorithms, under va-
riety of different sensor configurations, measurement noises and the uncertainties,
showed that identification and estimation achieved by time-varying-gain MMAE al-
gorithms show only slight improvements over constant-gain MMAE algorithms during
the initial transient. Thus, it is practical to use the steady-state KFs.
Information methods have been suggested by many authors for the solution of the
related problems of hypothesis testing, signal selection, and model identification. The
Kullback information metric [100] has proved to be useful in the analysis of parameter
estimation and model identification technique. It was employed by many people in
their studies of parameter estimates given stationary Gaussian observations, e.g. [16].
In this thesis we use the definitions and information measures employed by Baram [44],
which proved to possess valuable properties lacked by the Kullback information mea-
sure, such as the metric property on the parameter space; see [44,45,52,98,99] for more
details. In Ref. [16, ch. 10] there are similar proofs but in the framework of Kullback
Information function as well as evaluating power spectra of the signal model.
The analysis presented in this section is fairly general, applying to any model set
M, finite or nonfinite, and requires no assumptions of Gaussianess or stationarity.
59
where α can now be any element of T ≡ M ∪ {∗} and where ∗ denotes the “true”
parameter. If for some pair of parameters α1 , α2 we have
then it is natural to say that the information in the observations contained in yt favors
α1 over α2 . The above equation is equivalent to
or
p(yt ; α1 )
ln >0
p(yt ; α2 )
p(yt ;α1 )
Therefore, ln p(y t ;α2 )
can be regarded as a measure of information in yt favoring α1
over α2 . Similarly,
i) For any α ∈ M,
Jt (∗; α) ≥ 0 (3.45)
60
i.e., on the average, the true model is always favored by the observations.
i.e.
dt (αi ; αj ) = Jt (αi ; αj ) (3.46)
where x̃i (t|t) is any other causal estimate. Assuming knowledge of all matrices
in the model, the solution to this problem is well known (e.g. [56]) and is given by
utilizing a bank of discrete-time Kalman filters.
Returning now to the linear-Gaussian case with stationary system and model, con-
sider the true system described by equations of the form eq. (3.77) through eq. (3.86)
assuming u(t) ≡ 0 (i.e. the true plant is due to the stochastic inputs only). This true
system can then be specified by the N dimensional time invariant matrices:
denotes the predicted observation error covariance assuming that α is the true pa-
rameter. If each model in eq. (3.48) is observable and controllable (see [56]) the
steady-state limit
S(α) = lim S(t; α) (3.50)
t→∞
1 1
E∗ {ζ(yt |yt−1 ; α)} = lndet[S(α)] + tr[S −1 (α) S∗ (α)] (3.54)
2 2
62
for all α ∈ M.
This means that maximum likelihood estimates converge to the parameter in M
closest to the true model. This also represents a generalization of the consistency
result shown in Ref. [99]. Indeed, in view of eq. (3.53) – eq. (3.55), condition (3.56)
holds if and only if
and the same identifiability issues as those discussed in [99] are relevant here to α0
which satisfies eq. (3.56). In other words, the consistency and identifiability properties
connected with the true parameter in [99] generalize here to the parameter which
minimizes the information distance to the true parameter.
Extension of this section that leads to the Baram Proximity Measure is presented
in the next section.
this thesis. Baram in Ref. [44] has also studied the case when deterministic part also
exists but the only convergence proofs are for the case in the presence of stochastic
inputs but with deterministic control signal u=0. The stochastic inputs are assumed
to provide sufficient persistent excitation to the plant. We also will use the latter
case in the RMMAC methodology in the closed-loop case which also considers the
feedback control u(t).
As discussed in Section 3.3, the MMAE method was developed with the implicit
assumption that one member of the set of hypothesized models exactly matches the
true system. Under this assumption, Baram [44] has shown that the identification
algorithm (i.e., the value of the probability when all feedback gains are set to 0) con-
verges to the matched model when the input is ergodic. This convergence was verified
by using both maximum likelihood and related Bayesian estimates for general obser-
vation sequences. He also showed that when deterministic inputs are used and none
of the models match the true system the convergence properties are indeterminate -
the model to which the probability converges is a function of the input. Furthermore,
his results require the ergodicity of the residual which clearly is not always guaranteed
when the probabilistically weighted control is applied. Thus, no general convergence
result has been derived for the closed-loop adaptive situation.
In this section, we restrict our attention to linear systems driven by white Gaussian
inputs having time-invariant statistic. We make the assumption that the system has
attained steady-state, i.e. that all signals of interest are stationary. We will see that
the Kalman filtering theory reviewed in Section 3.2 plays a key role in the subsequent
development.
Consider the discrete (unknown) system
xt+1 = A∗ xt + L∗ ξt
(3.58)
yt = C∗ xt + θt
where ∗ denote the true system, {ξt } and {θt } are zero-mean white Gaussian
sequences, mutually uncorrelated and uncorrelated with x0 with
Based on a bank of N-discrete KFs, a set of models for the system can be denoted
by
Let ŷj,t denote the one-step prediction of yt , given the past observation Y (t − 1),
and assume that the j-th model is the true one. For each i, j ∈ {∗ ∪ K} let
denote the prediction error covariance matrices according to the respective mea-
sure and
1
− 21 (yt −yj,t )T Σj −1 (yt −yj,t )
fj (yt |Yt−1 ) = [(2π)m |Σj |]− 2 e
Both ŷj,t and Σj , are generated in essence by the discrete KF#k corresponding
to model Mj .
For notational purposes, assume that index ’i’ shows the true parameter. The
dynamic equation generating simultaneously the state xt and its one-step prediction
by the k-th KF, x̂j,t , is
where Kj is the constant KF gain matrix corresponding to the model Mj for all
k ∈ M where
T T
Ψij,t+1 = Aij Ψij,t Aij + Lij Qi Lij (3.61)
Then let
Ψij = lim Ψij (t, t1 ) (3.62)
t1 →−∞
Finally we have
T
Γij = Cji Ψij Cji + Ri (3.63)
It is well known that the limit of eq. (3.61) exists and is finite if the matrix Aij
has all its eigenvalues inside the unit circle, i.e. the spectral radius satisfies
This is the case if for each j ∈ {∗ ∪ M} Aj has all its eigenvalues inside the
unit circle and (Aj , Cj ) is observable. Note, however, that these conditions are only
sufficient, not necessary, for Γij to be finite, since eq. (3.63) may be finite even if Ψij ,
obtained as the limit value of eq. (3.61), is not finite.
66
An alternative way to solve eq. (3.61) is as follows: If the assumption of eq. (3.64)
is satisfied, the steady-state solution of eq. (3.61) can be found as follows.
At the steady-state,
T T
Ψij,t = Aij Ψij,t Aij + Lij Qi Lij (3.66)
Eq. (3.66) is the discrete Lyapunov matrix equation and can be solved by Matlab.
It can also be analytically solved by using the two following definitions.
Definition 3.5.2. The vec operator creates a column vector from a matrix A by
stacking the column vectors of A = [a1 a2 . . . at ] below one another:
a1
a
2
vec(A) = . (3.68)
..
at
x1
P
m x
2
(AXB):,k = AXbk = A xi bi,k = [b1,k A b2,k A ... bm,k A] .
i=1 ..
xt
| {z }
vec(X)
= ([b1,k b2,k ... bm,k ] ⊗ A])vec(X)
| {z }
bT
k
Stacking the columns
together
(AXB):,1 bT1 ⊗ A
(AXB) bT ⊗ A
:,2 2
vec(AXB) = .. = . vec(X) = (BT ⊗ A)vec(X)
. .
.
(AXB):,n bTn ⊗ A
Corollary. vec(AB) = (I ⊗ A)vec(B) = (BT ⊗ I)vec(A)
By taking vec operator for both sides of eq. (3.66) and simplifying we have
h i
i i iT
Ψij,t i i −1
= unvec (I − Aj ⊗ Aj ) vec(Lj Q Lj ) (3.69)
where dij is defined as the Baram Proximity Measure (BPM) of the k-th filter
generated by [44, eq. (5.16)]
© ª
dij = ln |Σj | + tr Σj −1 Γij (3.71)
where Ψij is the steady-state solution of eq. (3.61) (or equally of eq. 3.69).
68
P (T −1 A) = P (A)
for all A ∈ U.
2. A stochastic sequence {xt } on (Ω, U, P ) is said to converge almost everywhere
(a.e.) or almost surely (a.s.) to a random variable x on (Ω, U, P ) if
lim xt = x a.e.
t→∞
example in communication and signal theory, in the study of queues, or in time series
coming from economic and physical problems. In particular, stationary processes
arise as long-time limits of random dynamical systems, after transient phenomenon
due to the influence of initial conditions die away.
Mathematically, let {xt } be a stochastic sequence on (Ω, U, P ) with values in
(R` , B ` ), where (R` ) is the l-dimensional Euclidean space and (B ` ) is the σ-algebra
` `
of Borel sets of Rl ; see Ref. [103, ch. 3]. Let B∞ be the σ-algebra of Borel sets of R∞
`
where R∞ = R` × R` × . . .. Then {xt } is said to be stationary if for each k ≥ 1
`
for every C ∈ B∞ .
5. A stationary sequences {xt } on (Ω, U, P ) is said to be ergodic if every invariant
event in U has probability zero or one. It can be shown (see e.g. [101]) that {xt } is
generated by a measure preserving transformation T (the shift operator), i.e.
Let {xt } be a vector valued stochastic sequence from (Ω, U, P ) into (R` , B ` ) such
that the probability density with respect to the Lebesgue measure on (R` , B ` ) of {xt }
is Gaussian on R` , with
and
© ª
E (xt − mx )(xt − mx )T depends only on k.
Proposition 3.5.2. (Ergodicity condition: see [104, pp. 257-260] and [101, pp. 494])
where
T
<(k) = E{rt rt+k } (3.74)
is called the covariance function and |<(k)| denotes the determinant of the matrix
(
© ª Cji Ψij CjiT + Ri ; k=0
<(k) = E (yt − ŷj,t+k )(yt − ŷj,t+k )T = (3.75)
Cji Ψij (Aij )k CjiT ; k>0
¯ ¯ ¯ ¯ ¯ ¯k
|<(k)| = ¯Cji CjiT ¯ ¯Ψij ¯ ¯Aij ¯
¯ ¯
since all eigenvalues of Aij are inside the unit circle, i.e ¯Aij ¯ < 1, hence
t
X t
X
¯ i iT ¯2 ¯ i ¯2 ¯ i ¯2k
lim 2 ¯
|<(k)| = Cj Cj ¯ ¯ ¯
Ψj lim ¯Aj ¯
t→∞ t→∞
k=1 k=1
¯ i iT ¯2 ¯ i ¯2 ¯ i ¯2
¯Cj Cj ¯ ¯Ψj ¯ ¯Aj ¯
= ¯ ¯2 <∞ (3.76)
1 − ¯Ai ¯
j
It yields
t
1 X
lim |<(k)|2 = 0
t→∞ t + 1
k=0
3.5.2 Formulation
As mentioned above, this section deals only with linear state-space models. Our
notation is as follows.
Discrete State Dynamics of Plant:
Measurement Equation:
E{ξ(t)} = 0 (3.81)
cov[ξ(t), ξ(τ )] = Qδtτ (3.82)
Q = QT ≥ 0 (3.83)
E{θ(t)} = 0 (3.84)
cov[θ(t), θ(τ )] = Rδtτ (3.85)
R = RT > 0 (3.86)
(i.e., every measurement is corrupted by white noise) . Also, x(0), ξ(t), θ(τ ) are
assumed mutually independent for all t, τ .
It was assumed so far that the system under study has been modeled in the
above form, but that some parameters still are uncertain and need to be estimated.
Typically those would be coefficients in the entries of the model matrices in the
state-space format. Note first that if we denote those unknown parameters by the
vector α, the dependency of the model matrices can be made explicit by the notation
A(α), B(α), etc. Note also that the system can be time varying but that α must be
time invariant (at least according to the time scale of the identification experiment).
72
All stationarity and ergodicity assumptions required to prove the posterior probability
convergence results must hold. These convergence conditions are obtained in terms
of the second order statistics associated with the models in the model set. We shall
recall the as “theoretical assumptions” throughout this thesis.
Condition c1. For each j ∈ {∗ ∪ M}, the residual covariance Σj exists and has a
finite positive definite value; as in eq. (3.50).
Condition c2. For each j ∈ {∗ ∪ M} the residual sequence yt − ŷj,t is ergodic
satisfying the ergodicity condition of eq. (3.73).
Condition c3. There exists some parameter k ∈ M such that
In this section, a brief summary is presented on how to robustify the design of the
KFs in the RMMAC architecture using the BPM, so as to ensure “correct model
identification” by the posterior probabilities. We stress that a KF is optimal if indeed
the data is generated by the same model as that used to design the KF. In adaptive
control, however, the true parameter is NOT identical to that of the “closest” KF in
the sense of Section 4.3. Also, the unmodeled dynamics are not taken into account
in the design of the KF. Moreover, the KFs and the PPE must perform well even
when some of the stochastic assumptions are violated. Thus, we must “robustify” the
KF to be tolerant of such “errors”. The time-honored engineering practice of using
suitable “fake plant white noise” is a very useful tool. It is known, see e.g. [95], that
judicious use of “fake plant white noise”, which causes the KF gains to increase and
pay more attention to the measurements, can be a very valuable tool in improving
the quality and speed of probability convergence. We believe that future fundamental
73
The ergodicity condition given in eq. (3.73) should be checked for all values of
the parametric uncertainties of the unknown plant. If the ergodicity conditions are
not satisfied for some parameters of the model, the assumptions of the convergence
proofs are violated and there should not be any expectation of model probabilities to
convergence, unless one can find a “fix”.
One should think of zero as the precision of numerical representation of any com-
puter’s roundoff limitations. What was important to our numerical simulations was
not the quantity of the right-side of eq. (3.73) for each model alone, but the relative
quantities of all models were important. For example, if the ergodicity conditions of
all models are, say about 10−12 , one can regard them as zero as they will satisfy the
condition. But if this quantity is, say for Model#i, 10−16 and for the other models,
say, about 10−10 , it means that the ergodicity condition is violated and one has to fix
it before applying MMAE and/or RMMAC.
One way to overcome this problem is to monitor the signals of residuals and then
add an appropriate “plant fake noise” in those measurement which do not include
any plant disturbance uncertainty. If so, the uncertainty of the plant parameters are
propagated in the KFs in order to robustify them to the parametric uncertainties.
This trick is used in fact to satisfy the theoretical assumptions and the ergodicity
condition. Otherwise, the convergence of the probabilities may not occur. The ap-
propriate fake disturbance intensity (Ξf ) design process can involve some trial and
error. Same more details can be found in Section 8.3.
We shall also note that in Ref. [95] relating to the problems of “global positioning
systems, inertial navigation, and integration” the authors discuss extensively the need
for using “fake white noise” in real engineering applications.
74
4.1 Introduction
Modern control theory using H∞ and µ-synthesis techniques represents a new and
powerful approach to multivariable feedback control system design. Multiple uncer-
tainty sources at different locations in the plant, external disturbances as well as
performance specifications are considered when designing for robust performance.
If multiple uncertainty sources are present in the system, H∞ compensators will
be too conservative since they do not account for the structure of the uncertainty.
The structured singular value, µ, is a powerful tool introduced in [105]. The so-called
complex µ-synthesis diminishes this problem but is still conservative in the sense
that only complex-valued uncertainties are dealt with. If, in addition, we have real
parameter uncertainties, complex-µ synthesis will be even more conservative, while
mixed-µ synthesis removes this conservatism to a significant degree.
It is emphasized that no original material is presented in this chapter. Simply,
our summary of µ-synthesis is necessary to better understand the proposed RMMAC
design methodology.
In this chapter, the general method of formulating control problems is reviewed.
Within this framework, we recall the general method for representing uncertainty for
multivariable systems. We see that, by using µ, it is possible to get “nonconservative
75
76
sufficient conditions” for robust stability and also for robust performance. In fact, the
“optimal” robust compensator can be designed using the complex-µ synthesis (also
called DK-iteration) or the mixed-µ synthesis (also called DGK-iteration) which both
involve solving a sequence of optimized scaled H∞ problems.
This chapter is organized as follows. Section 4.2 reviews the literature on robust
control including an outline of uncertainty models, robustness in control systems,
and H∞ -optimal design. In Section 4.3 we see how the structured singular value
can be used to synthesize a µ-“optimal” robust compensator in the presence of both
complex-valued and real-valued parameter uncertainties. Some technical issues are
summarized in Section 4.4.
Stability and performance are two of the fundamental issues in the analysis, design,
and evaluation of control systems [106–108]. Stability means that, in the absence of
external excitation, all signals in the system decay to zero. Alternatively, one can
define bounded-input bounded-output (BIBO) stability. Stability of the closed-loop
system is an absolute requirement since its absence causes signals to grow with-
out bound, eventually destroying the plant. This is what happens when an aircraft
crashes, or a satellite spins out of control or a nuclear reactor core heats up uncon-
trollably and melts down. In many interesting applications the open loop plant is
unstable and the objective of feedback control is, of course, to stabilize the system.
77
While feedback is necessary to make the system track the reference inputs and reject
plant disturbances, its presence in control systems causes the potential for instability
to be ever present and very real. In engineering systems it is of fundamental impor-
tance that control systems be designed so that stability is preserved in the face of
various classes of uncertainties. This property is referred to as robust-stability.
In most control systems the compensator remains fixed during operation while
plant parameters may slowly vary over a wide range about their nominal values. The
term robust parametric stability refers to the ability of a control system to main-
tain stability despite such large variations. Robustness with respect to parameter
variations is necessary because of inherent uncertainties in the modeling process and
because of actual parameter variations that occur during the operation of the system.
The performance of a system usually refers to its ability to closely track reference
inputs and reject exogenous disturbances. A well designed feedback control system
should be capable of tracking reference signals, belonging to a class, without exces-
sive error and despite various types of uncertainties. In other words, the worst-case
performance over the uncertainty set should be acceptable. This is, roughly speaking,
referred to as robust-performance.
In summary, the requirements of robust-stability and robust-performance are
meant to ensure that the control system functions reliably, despite the presence of
significant uncertainties regarding the model of the system and the precise descrip-
tion of the exogenous signals to be tracked or rejected. In the sequel we discus these
requirements and their impact on control system design in some detail.
The uncertainties encountered in control systems are both in the environment and
within the system. In the first place there are disturbance signals tending to drive
the system in an undesired manner. For example, the load torque on the shaft of
an electric motor, whose speed is to be maintained constant, can be regarded as a
disturbance.
As far as the system is concerned the main source of uncertainty is the behavior
78
of the plant. These uncertainties can occur, for example, due to changes of operating
conditions. as in an aircraft flying at various altitudes and speeds, or a power system
delivering power at differing load levels. Large changes can also occur in an uncon-
trolled fashion, for example when sensor or actuator failures occur. The complexity of
even the simplest plants is such that any mathematical representation of the system
must include the inevitable presence of significant uncertainties.
In analysis and design it is customary to work with a nominal mathematical plant
model. This is almost always assumed to be linear and time invariant (LTI), because
this is the only class of systems for which there exists any reasonably general design
theory. Nevertheless, such LTI models are usually a gross oversimplification and
it is, therefore, necessary to test the validity of any proposed design by testing its
performance when the actual or true model is significantly different from the nominal
one.
The LTI models that are usually employed are approximations which are made
to render the design and analysis of complex systems tractable. In reality, most
systems are nonlinear and a LTI model is obtained by fixing the operating point
and linearizing the system equations about it. As the operating point changes so do
the parameters of the corresponding linear approximation. Thus, there is significant
uncertainty regarding the “true” plant model, and it is necessary that a compensator
that stabilizes the system does so for the entire range of expected variations in the
plant parameters. In addition, other reasonable, but less structured uncertainties
of the plant model must also be tolerated without disrupting closed-loop stability.
These unstructured uncertainties arise typically from reducing the order of a complex
model, possibly infinite-dimensional, by retaining only some of the dominant modes,
which usually lie in the low-frequency range. Therefore unstructured uncertainties,
or dynamic modeling errors, are typically significant in the high-frequency region.
We know that precise knowledge of the plant is required for a proper design of a
compensator K(s), otherwise the compensator is bound to fail when driving the real
system. Nevertheless, an exact knowledge of the plant is not always possible: models
may be inaccurate and may not reflect the changes suffered by the plant with time.
Therefore, it is often assumed that the real plant, denoted by P , is unknown but
79
P = P0 + ∆A (4.1)
or multiplicative as
P = (I + ∆M )P0 or P = P0 (I + ∆M ) (4.2)
Here, P0 denote the nominal plant model, which is assumed to be linear and finite
dimensional.
80
The control problem here is to ensure that the closed-loop remains stable under
all such uncertainties, and the worst-case performance is acceptable in some precise
sense.
Both additive ans multiplicative uncertainties can be characterized as norm bounded
quantities which are bounds on the norms of operators corresponding to systems, i.e.,
using the H∞ norm.
Let P ∈ P be any member of the set of possible plants P and let P0 ∈ P is
the nominal model of the plant. The magnitude of the actual uncertainty L may be
measured in terms of a bound on the maximum singular value σ̄(L),
σ̄(L) ≤ w(ω) ∀ω
where
w(ω) = max σ̄(L)
P ∈P
The bound w(ω) can also be interpreted as a scalar weight on a normalized un-
certainty ∆(s),
Sometimes, matrix weights are used to describe the uncertainty, i.e. L = W1 ∆W2 .
Nevertheless, neither the scalar weight w(s) nor the transfer function matrices W1
and W2 generally constitute an exact description of the uncertainty. This leads to a
conservative uncertainty description as discussed below.
σ̄(∆i ) ≤ 1 ∀ω
and weighting matrices W1i and W2i are used to describe the actual uncertainty
Li
Li = W2i ∆i W1i
81
The individual uncertainties ∆i are combined into a single (large) block diagonal
uncertainty matrix
∆ = diag{∆1 , . . . , ∆m } (4.3)
w z
P
u y
K
The compensator design problem is divided into the analysis and the synthesis
phases. The compensator K is synthesized such that some measure, i.e. a norm, of the
82
transfer function from w to z is minimized, e.g. the H∞ norm. Then the compensator
design problem is to find a compensator K – that generates a signal u considering the
information from y to mitigate the effects of w on z – minimizing the closed-loop H∞
norm from w to z. For the analysis phase, the scheme in Figure 4.1 is to be modified
to group the generalized plant P and the resulting synthesized compensator K in
order to test the closed-loop performance achieved with K. How to group P and K
is treated later on. To get meaningful compensator synthesis problems, weights on
the exogenous inputs w and outputs z are incorporated. The weighting matrices are
usually frequency dependent and typically selected such that the weighed signals are
of unit magnitude, i.e. the norm from w to z should be less than 1, if possible.
A weighted generalized plant P has been derived to allow synthesizing the com-
pensator K. If the generalized plant P is partitioned as
P11 P12
P = (4.4)
P21 P22
such that its parts are compatible with the signals w, z, u and y in the generalized
control configuration, then
z = P11 w + P12 u
(4.5)
y = P21 w + P22 u
In order to obtain a good design for K, a precise knowledge of the plant is required.
The dynamics of interest are modeled but this model may be inaccurate and may not
reflect the changes suffered by the plant with time. To deal with this problem, the
concept of model uncertainty must be considered. The unknown plant P is assumed
to belong to a “legal” class of models, P, built around a nominal model P0 . The set of
models P is characterized by a matrix ∆, which can be either a full matrix or a block
diagonal matrix that includes all possible system structured uncertainties. We also
use the weighting matrices (and incorporate them into P ) to express the uncertainty
in terms of normalized uncertainties in such a way that k∆k∞ ≤ 1.
Throughout this chapter (and in the next chapters), ∆ indicates the normalized
model uncertainty block, z is a signal we would like to keep small, and w repre-
sents external disturbances. The general control configuration in Figure 4.1 may be
extended to include model uncertainty as shown in Figure 4.2.
D
q p
w z
P
u y
∆
w1 z1
w2 M z2
To represent any control problem with uncertainty by the general control con-
figuration in Figure 4.2 it is necessary to represent each source of uncertainty by
a single uncertainty block ∆, normalized such that k∆k∞ ≤ 1. The uncertainties
can represent parametric uncertainty, unmodeled dynamics, etc. followed in detail in
Section 4.2.
In summary, the set of models P is characterized by a matrix ∆ which can be
either a full matrix or a block diagonal matrix including all possible uncertainties
representing uncertainty to the system. In this section, we present the general frame-
work for the analysis and synthesis of compensators for robust stability and robust
performance with MIMO systems. The following terminology is used:
Definition 4.2.1. The closed-loop system has Nominal Stability (NS) if the com-
pensator K internally stabilizes the nominal model P0 .
Definition 4.2.2. The closed-loop system has Nominal Performance (NP) if the
performance objectives are satisfied for the nominal model P0 .
Definition 4.2.3. The closed-loop system has Robust Stability (RS) if the compen-
sator K internally stabilizes every plant P ∈ P, i.e., in Figure 4.3, Fu (M, ∆) is stable
∀∆, k∆k∞ ≤ 1.
85
Definition 4.2.4. The closed-loop feedback system has Robust Performance (RP) if
the performance objectives are satisfied for P ∈ P, i.e., in Figure 4.3,
In this section, we present a brief sketch of H∞ optimal control theory and its connec-
tion to robust control; see, e.g., [61]. The state-of-the-art non-adaptive robust multi-
variable feedback control system design is to synthesize a control law which maintains
system response and error signals to within prespecified tolerances despite the effects
of uncertainty on the system. Uncertainty may take many forms (see Section 4.2.2)
but among the most significant are noise/disturbance signals and uncertainty in the
multivariable differential equations that model the plant (or transfer function model-
ing errors). In such applications, design by well-known H∞ -optimization method has
been considered as a powerful design method which involves the minimization of the
peak magnitude of a suitable closed-loop system function [63, 106, 110]. It can also
be very well suited to frequency response loop shaping. Moreover, robustness against
plant uncertainty in H∞ methodology may be handled more directly than with the
classical LQG (i.e. H2 -optimization) theory [61, 71].
An important application of the H∞ control problem arises when studying ro-
bustness against model uncertainties. It turns out that the condition that a control
system is robustly stable in spite of a certain kind of model uncertainties can be
expressed quantitatively in terms of an H∞ norm bound which the control system
should satisfy. The connection of the H∞ norm and robust stability will be described
in this section where we present a state-space solution to the H∞ -optimal control
problem.
As in Section 4.2.3, we consider the control system in Figure 4.1. The performance
measure to be minimized is now taken as the H∞ -norm of the closed-loop transfer
function, i.e., we consider the cost
The control problem is most conveniently solved in the time domain using a state-
space based approach, see e.g. [61, 63]. We will assume that the plant P has the
state-space representation
Notice that the only inputs to the system are w and u which will be defined for
t ≥ 0 only.
Assumptions on the system matrices are as follows.
T
(A1) D12 D12 is nonsingular,
T
(A2) D21 D21 is nonsingular,
The direct minimization of the cost J∞ (K) turns out to be a very hard problem,
and it is therefore not feasible to tackle it directly. Instead, it is much easier to
construct conditions which state whether there exists a stabilizing compensator which
achieves the H∞ -norm bound
J∞ (K) < γ (4.10)
for a given γ > 0. In that case, the conditions also provide a specific compensator
which achieves the bound eq. (4.10). One can then use the conditions for checking
the achievability of eq. (4.10) for various values of γ, and in this way determine
the minimum of J∞ (K) to any degree of accuracy. Such a procedure is called “γ-
iteration”.
In order to derive conditions for checking whether there exists a compensator
which achieves the bound eq. (4.10), recall that the H∞ performance measure can be
characterized in terms of the worst-case gain in terms of L2 -norm, i.e.,
½ ¾
kzk2
J∞ (K) = sup : w 6= 0 (4.11)
kwk2
87
kzk2
< γ, all w 6= 0 (4.12)
kwk2
or
L(w, u) = kzk22 − γ 2 kwk22 < 0, all w 6= 0 (5.6) (4.13)
As the H∞ -optimal compensator which achieves the bound eq. (4.13) will be
derived in the time domain, it is convenient to give an explicit time-domain expression
of the inequality (4.13). By Parseval’s theorem and the time-domain expression of
the L2 -norm we have that eq. (4.13) is equivalent to
Z∞
£ ¤
L(w, u) = z(t)T z(t) − γ 2 w(t)T w(t) dt < 0, all w 6= 0 (4.14)
0
The problem has thus been stated in the form of a particular dynamic game
problem. Here the first “player” w tries to make the cost L(w, u) as large as possible,
while the second “player” K attempts to minimize the cost L(w, u) < 0 regardless of
the action of w.
As the system is linear and the cost L(w, u) is quadratic, the associated dynamic
game problem is a so-called linear quadratic game problem, in analogy with linear
quadratic control.
The solution of the H∞ -optimal control problem via the game problem defined by
eq. (4.14) has a similar structure as the solution of the H2 problem; see [61]. Thus,
the solution is obtained in two stages: the first stage consists of an H∞ -optimal state-
feedback control problem and an associated variable transformation, and the second
stage consists of an H∞ -optimal estimation problem.
The aim is then to find an H∞ -optimal feedback law which achieves the inequality
J∞ (Kx ) < γ, provided such a compensator exists.
88
Consider the system eq. (4.9). Suppose that the assumptions (A1) to (A4) hold.
Assume that the control signal u(t) has access to the present and past values of the
state, y(τ ), τ < t. Then there exists a H∞ compensator such that J∞ (K) < γ, i.e.,
the inequality (4.14) holds for all w 6= 0, if the following conditions are verified. Note
that if any one is violated γ must be increased.
1- There exists a positive (semi) definite solution matrix X to the Modified Control
Algebraic Riccati Equation (MCARE)
AT X + XA − X[B2 (D12
T
D12 )−1 B2T + γ −2 B1 B1T ]X + C1T C1 = 0 (4.16)
2- The matrix
T
A − B2 (D12 D12 )−1 B2T X + γ −2 B1 B1T X
is strictly stable, i.e. all its eigenvalues have negative real parts.
3- There exists a positive (semi) definite solution matrix Y to the Modified Filter
Algebraic Riccati Equation (MFARE)
4- The matrix
A − Y C2T (D21 D21
T −1
) C2 X + γ −2 Y C1T C1
is strictly stable.
5- The spectral radius must satisfy the inequality
where
ρ(A) ≡ max |λi (A)|
i
where r(s) is a real, rational, stable, and proper (RRSP) transfer function. Under
the assumption that every admissible P (s) and P0 (s) have the same number of unsta-
ble poles it can be easily shown, based on the Nyquist criterion that, K(s) stabilizes
the entire family of plants if and only if
° ¡ ¢ °
°K(jω) I + P0 (jω)K(jω) −1° |r(jω)| < 1, for all ω ∈ R (4.20)
∞
The robust stability question posed above can also be formulated in terms of
uncertainties which are H∞ functions and lie in a prescribed ball
The conditions in eqs. (4.20), (4.21), and (4.22) can all be derived from the Nyquist
criterion by imposing the requirement that the number of encirclements required for
stability of the nominal system remain invariant under uncertainties. This amounts
to verifying that
|I + P (jω)K(jω)| =
6 0, for all ω ∈ R (4.23)
90
|I + K(jω)P (jω)| =
6 0, for all ω ∈ R (4.24)
for deriving eq. (4.22). The conditions of eqs. (4.21) and (4.22) can be derived
from a general result called the Small Gain Theorem as discussed in the sequel.
The uncertain closed-loop transfer function from w to z in Figure 4.3 was related
by the upper LFT
Assuming that the system is nominally stable, the only source of instability in the
upper LFT eq. (4.26) is the term (I − M∆)−1 . Therefore, the stability of the system
in Figure 4.3 is equivalent to the stability of the structure shown in Figure 4.4, where
M = N11 , i.e. the portion of the transfer matrix function N seen by the uncertainty
block ∆.
The system shown in Figure 4.4 is stable if and only if det(I − M∆) does not
encircle the origin as s crosses the Nyquist D contour for all possible ∆. Because the
uncertainties are norm bounded, i.e., σ̄(∆) ≤ 1, this is equivalent to [111]
∆
q p
This condition is not itself useful to check the stability of the M∆ structure since it
must be tested for all possible uncertainties ∆. What is desired is a condition on the
matrix M, preferably on some norm of M. The “small gain theorem” provides the
tool to analyze the stability robustness of this model establishing a condition on M
so that it can not be destabilized by ∆.
Theorem 4.2.1. (Small Gain Theorem) Assume that both M(s) and ∆(s) are stable.
Then the interconnected M∆ system shown in Figure 4.4 is well-posed and internally
stable for all uncertainties ∆(s) with σ̄(∆) ≤ 1 if and only if
saw above, the stability of the N ∆ structure is determined by the term (I − N ∆)−1 .
Therefore, to test the stability of the structure in Figure 4.3 is equivalent to test the
stability of the M∆ structure shown in Figure 4.4, where M = N11 .
By assuming norm bounds on each individual uncertainty, e.g., σ̄(∆i ) ≤ 1 it is
possible to derive a necessary and sufficient non-conservative condition on M so that
it can not be destabilized by ∆. This is obtained by the structured singular value.
singular values, σ̄, and the spectral radius, ρ. The SSV is used to get necessary
and sufficient, non-conservative, conditions for robust stability and also for robust
performance in the presence of structured uncertainties. The SSV is defined to obtain
the tightest possible bound on M such that det(I − M∆) 6= 0.
In Section 4.2.2 the concept of model uncertainty was presented, assuming that
the real plant was unknown but belonging to a family of models built around the
nominal plant. Also, in Section 4.2.3 the general configuration to formulate control
problems was introduced. In this section, the structured singular value (µ) synthesis is
outlined in two uncertainty cases, complex and mixed, to get necessary and sufficient
conditions for stability- and performance robustness.
Definition 4.3.1. For a square complex matrix M, the Structured Singular Value,
µ∆ (M), is defined for some structured ∆ at each frequency such that
1
µ∆ (M) = (4.28)
inf {σ̄(∆) : det(I − M∆) = 0}
∆
© ri ×ri
ª
∆s = ∆ = diag(∆1 , . . . , ∆s )|∆i ∈ H∞ , k∆i k 6 1 (4.30)
ri ×ri
where H∞ denotes the set of stable ri × ri transfer functions with bounded H∞
norm. The uncertainty set ∆s is thus characterized by the following set of parameters:
the number of blocks s and the dimensions of the uncertainty blocks, ri . Notice that
the structure of ∆ ∈ ∆s implies that k∆k∞ ≤ 1 as well.
The structured singular value turns out to be very hard to calculate numerically
and no efficient algorithm for its computation exists. Instead, a more tractable ap-
proach is to calculate an upper bound on µ which is easier to compute. In particular,
define the set of diagonal matrices D with the following structure corresponding to
the structure of ∆s
Ds = {D = diag(d1 I1 , . . . , ds Is )} (4.31)
where Ir denotes the r × r identity matrix, and the di are complex-valued scalar
constants.
To be more specific, the µ-upper bound is derived by the computation of non-
negative scaling matrices Dl and Dr defined within a set D that commutes with the
structure ∆. For a detailed discussion on the specification of such a set D of scaling
matrices, see e.g. [13]. The commutation of D with ∆ implies that Dr ∆ = ∆Dl and
µ∆ (M) = µ∆ (Dl M Dr−1 ) for all Dl , Dr ∈ D. Then, it can be shown that the upper
95
w z
Dl-1 Dr
P
PD
u y
K
The optimization problem of eq. (4.32) is convex in D, i.e. it has only one mini-
mum. It has been shown that the inequality is, in fact, an equality if the number of
blocks in ∆ are less than 4 [3]. Otherwise, the upper bound seems to be tight (within
a few percent [13]). With these scaling matrices, the plant can be augmented as
illustrated in Figure 4.5. The resulting system PD fits into the standard H∞ optimal
control framework.
where Fu (N , ∆) is the upper LFT derived from the general block diagram for
analysis with uncertainty in Figure 4.3.
In order to evaluate robust performance using the structured singular value func-
tion, let us compare the condition in eq. (4.27) for robust stability and the formally
identical condition in eq. (4.33) for robust performance. From the small gain theorem
96
(Theorem 4.2.1) we know that stability of the M∆ structure in Figure 4.4, with a
full complex matrix ∆, is equivalent to kMk∞ < 1.
From this theorem we can conclude that the condition for robust performance
kFu k∞ < 1 is equivalent to the stability of the Fu ∆P structure, with ∆P being a
full complex matrix of appropriate dimensions.
The structure of ∆ is now given by
(" # )
∆ 0
∆= : ∆, ∆P ∈ RH∞ ; σ̄(∆) ≤ 1, σ̄(∆P ) ≤ 1 (4.34)
0 ∆P
Figure 4.6 shows the equivalence between robust performance (RP) and robust
stability (RS). Condition (4.33) for RP is satisfied if and only if the system N is
robustly stable with respect to the block diagonal uncertainty ∆ defined in eq. (4.34).
∆P
∆
w1 z1
N
w2 z2
Figure 4.6: Block diagram for testing both robust stability and robust performance.
Theorem 4.3.2. The nominally stable system N in Figure 4.6 subject to the block
diagonal uncertainty ∆, σ̄(∆) ≤ 1 satisfies the robust performance condition
if and only if
µ∆ (N ) < 1, ∀ω (4.35)
where the complex-µ is computed with respect to the structure ∆ in eq. (4.34)
and ∆P is a full complex uncertainty matrix with appropriate dimensions reflecting
performance specifications.
97
holds. If γ < 1, K(s) is robustly stabilizing and the problem is solved. Otherwise, go
to Step 2.
Step 2. (D-step). With K fixed, calculate a new scaling matrix D by minimizing
σ̄[Dr (jω)M(jω)Dl (jω)−1 ] (approximately) with respect to D(jω). If the procedure
gives a matrix such that σ̄[Dr (jω)M(jω)Dl (jω)−1 ] < 1, all ω, then the compensator
K is robustly stabilizing, and the problem is solved. Otherwise, go to Step 3.
98
The complex µ-synthesis (D − K iteration) in its present form and as presently is im-
plemented in current numerical Matlab software package [3,113,114] has substantial
disadvantages, that is, when real parameter uncertainty is present in the system, the
resulting compensator can be arbitrarily conservative.
In this section, a summary of the background for dealing with combined real
and complex uncertainties, so called mixed-µ synthesis is presented. The theoretical
foundation for the mixed-µ analysis and synthesis is too copious for a comprehensive
review in the framework of this dissertation. Hence, only a summary is provided here.
For a more in-depth discussion the reader is referred to e.g. Refs. [46, 47, 110, 115]2 .
2
However, all our mixed-µ designs in the RMMAC have been carried out with advanced D,G-K
iteration software provided to us by Prof. G.J. Balas [114] of the University of Minnesota.
99
Consider the robust performance problem of the mixed-µ setup [116] depicted in
Figure 4.7.
δ1 I 0
δ2 I
0 0
0 δ pI
∆= ∆1 ( s) 0
∆ 2 ( s)
0 0
0 ∆ n ( s)
0 0 ∆ P (s )
q p
P(s )
w z
u y
K ( s) M ( P, K )
Figure 4.7: The mixed-µ setup. The uncertainty block, ∆, contains the following
blocks:
UNCERTAIN REAL PARAMETERS: All δi are real scalars, |δi | ≤ 1, ∀i.
Size of identity matrix, I, depends on the number of times the same uncer-
tain parameter exists.
COMPLEX MODEL ERRORS: All ∆j (s) reflect complex-valued unmod-
eled dynamics at different places, k∆j (s)k∞ ≤ 1, ∀j.
PERFORMANCE: The ∆P (s) block captures performance specifications
in the frequency-domain transformed into equivalent robustness problems,
k∆P (s)k∞ ≤ 1.
The definition of the mixed structured singular value (mixed-µ) depends upon the
underlying block structure of the uncertainties. The following types of uncertainties
are considered in this setup: repeated real scalars δ r , repeated complex scalars δ c ,
and full complex blocks ∆C . The non-negative integers mr , mc , and mC specify
the number of uncertainty blocks of each type while the positive integers ki ; i =
1, 2, . . . , mr + mc + mC , define the dimensions of each block (i.e., the multiplicity of
real or complex scalars should they be repeated, or the size of a full complex block,
100
δir ∈ R, δic ∈ C, ∆C
i ∈ C
kmr +mc +i ×kmr +mc +i
} (4.37)
where KS is the set of all real-rational, proper compensators that nominally sta-
bilize P . Unfortunately, the structured singular value cannot be calculated exactly.
Thus, µ(·) is typically approximated by its upper bound. Hopefully, the value of
µ and its upper bound are close. A lower bound on µ(·) can also be calculated to
judge the tightness of the upper bound. In order to formulate an upper bound on
the mixed-µ(M), the following sets of block diagonal frequency-dependent scaling
matrices are defined:
G = {diag(G1 , . . . , Gmr , 0Ikmr +1 , . . . , 0Ikmr +mc +mC ) : Gi = G∗i ∈ Cki ×ki } (4.41)
where 0 denotes a zero matrix and G∗ the complex conjugate transpose of G. The
upper bound on µ(M) is then given as [46]
½ µ µ ¶ ¶ ¾
¡ ¢ 1 1
2 −4 −1
¡ ¢ 1
2 −4
µ(M) ≤ inf inf β : σ̄ I +G DMD − jG I +G ≤1
D∈D β∈R β
P ∈G β>0
(4.42)
101
Now the µ-analysis problem in eq. (4.42) can be combined with eq. (4.39) to form
the mixed-µ synthesis problem:
where
µ µ ¶ ¶
¡ 2
¢− 14 1 −1
¡ 2
¢− 41
Γ = σ̄ I + G (ω) D(ω)M(jω)D (ω) − jG(ω) I + G (ω) .
β(ω)
(4.44)
If there is a β > 0, D ∈ D, and G ∈ G such that Γ ≤ 1 in eq. (4.44) then
µ∆ (M) ≤ β.
This bound, [46], is a derivative of an earlier bound in [115]. The smallest β > 0
for which D and G matrices exist which satisfy this constraint is what µ-Tools calls
the mixed-µ upper bound for M. Using manipulations that are now standard in
robust control theory, the computation of the best such β is reformulated into an
Affine Matrix Inequality (AMI) and solved with special purpose convex programming
techniques [114].
As in complex-µ synthesis, the problem of finding optimal “scales” D(ω), G(ω) and
β(ω) for a given compensator K is quasi-convex. Real-rational stable and minimum-
phase transfer matrices D(s) and G(s) are fitted to D(ω) and jG(ω), and augmented
to the plant in such a way that the resulting interconnection is stable. Holding D, G
and β fixed, the problem of finding K is reduced to a standard H∞ problem, which
is convex in the full order case. See Figure 4.8. This leads to the following D,G-K
iteration scheme for the mixed-µ synthesis problem:
D,G-K Iteration [47]:
1. Find initial estimates of the scaling matrices D(ω), G(ω) and the real positive
scalar β. One possibility for the scalings D(ω) is to choose them as the identity matrix
at each frequency point. If G(ω) is chosen to be the zero matrix at each frequency
point, then β is arbitrary (so choose, say, β = 1).
2. Fit state-space realizations D(s) and G(s) to the pointwise scaling matrices
D(ω) and jG(ω), so that D(jω) approximates D(ω), and G(jω) approximates jG(ω).
102
Now replace D and G with appropriate factors so that D, D−1 , Gh and GGh are all
stable, where Gh is a spectral factor satisfying (I + G∼ G)−1 = Gh G∼ ∼
h and G (s) =
GT (−s). Augment D and Gh with identity matrices, and G with a zero matrix, of
appropriate dimensions so that D, G, Gh are all compatible with P . Form the new
state-space system
PDG = (DP D−1 − β∗ G)Gh
where
µ³ ´− 14 µ 1 ¶³ ´− 41 ¶
2 −1 2
Γ = σ̄ I + G̃ (ω) D̃(ω)M(P, K)(jω)D̃ (ω) − j G̃(ω) I + G̃ (ω)
β(ω)
5. Find D̂(ω), Ĝ(ω) by solving the minimization problem
µ³ ´− 14 µ 1 ¶³ ´− 41 ¶
2 −1 2
inf σ̄ I + Ĝ (ω) D̂(ω)M(P, K)(jω)D̂ (ω) − j Ĝ(ω) I + Ĝ (ω)
D̂(ω)∈D β∗
Ĝ(ω)∈G
In order to augment the state-space realizations of D(s) and G(s) to the gener-
alized plant, a spectral factorization is used to determine a transfer function matrix
−1/2
for the expression (I + G2 (s)) :
Gh G∼ ∼
h = (I + G G)
−1
(4.46)
103
where G∼ (s) = GT (−s) (step 2). Assuming a minimal realization for the square
transfer function matrix P
A B
G∼ (4.47)
C D
R = I + DT D (4.48)
Q = I + DDT (4.49)
Accordingly,
−1 T −1 T − 12
A − BR D C − BR B X BR
P Gh ∼ (4.52)
−1 T − 12
Q (C − DB X) DR
- b *G
w z
Gh D -1 D
P
PDG
u y
K
Figure 4.8: Scaled system PDG for H∞ compensator design step in the mixed-µ (D,G-
K iteration).
4.4.1 Complexity of µ
As we saw in this chapter, the structured singular value measures the robustness of
uncertain systems and is in fact used to get necessary and sufficient, non-conservative,
conditions for robust stability and also for robust performance in the presence of
structured uncertainties.
Numerous researchers over the last decade have worked on developing efficient
methods for computing µ including the branch and bound algorithms that are sug-
gested to be used when there are a large number of uncertain parameters; see e.g.
[1, 13, 14, 117]. It is now known the general µ problem with mixed real/complex (or
pure real) uncertainty is a NP-hard problem [118–120]; Ref. [118] considerers the
complexity of calculating µ with mixed uncertainty in the framework of combinato-
rial complexity theory and also concludes that it is futile to pursue exact methods for
calculating µ for general systems with either pure real or mixed uncertainty for other
than “small problems”.
In the presence of high dimensional parametric uncertainties, i.e. by increasing
the number of uncertainties, not only the complexity of running the algorithms in
Matlab is increased, i.e. computation of the D,G-K iteration needs much more
time, but also it slows down the convergence. Moreover, if the number of parametric
uncertainties increases the multiplicity of possible local minima of this iteration will
increase. This exhibits another complexity of D,G-K iteration in the presence of high
105
The order of a system is the minimal number of states in its state-space realization
(also called the McMillan degree). The general model order reduction problem is as
follows [122].
Given an LTI system P of order n > r, find an LTI system P̂ = P̂ (s) of order
° °
° °
not greater than r such that °P − P̂ ° is as small as possible, where k·k is some norm
measure on the set of LTI systems. The norm is frequently a weighted H∞ norm,
° ° ° °
° ° ° °
such as °P − P̂ ° = °W −1 (P − P̂ )° .
∞
As we have remarked above, mixed-µ synthesis yields very high-order compen-
sators. These can be significantly simplified to ease implementations and simulations
efforts.
A significant disadvantage of both D − K or D,G-K iterations (complex or mixed-
µ techniques) is that the resulting compensator is of the same order as the generalized
plant. It is due to the fact that the frequency-dependent weights are included in the
design framework in order to achieve the desired performance characteristics and to
account for the structure in the uncertainty. Thus, the order of the generalized plant
is increased resulting in high order compensators. Also the dimensionality of the D
and G scales, see Figure 4.8, increase the order of the generalized plant and, hence,
of the H∞ compensator.
106
One possible solution to this problem could be to constrain the order of compen-
sator a priori in the design process. In this case, our designs throughout this thesis
showed that we have unacceptable performance. We do not recommend a priori
constraints on the order of compensator, and mainly on the D scales.
In defining the robust compensators throughout this thesis using the mixed-µ tool-
box of Matlab [114], we have applied the order reduction to the compensators. The
model reduction techniques that we used are all based on the Hankel singular values
of a system. They can achieve a reduced-order robust compensator that preserves the
major characteristics of the closed-loop system characteristics with minimal sacrifices
to robust performance. Other methods, such as the balanced truncation model re-
duction and state-space modal truncation/realization, resulted in very similar results
to the optimal Hankel norm model approximation that we used.
107
5.1 Introduction
In this chapter, a general philosophy and methodology is discussed for designing “ro-
bust” adaptive multivariable feedback control systems for linear time-invariant (LTI)
plants that include both unmodeled dynamics and a single uncertain real parameter in
the plant state-space description. The adjective “adaptive” refers to the fact that the
real parameter uncertainty and performance requirements require the implementa-
tion of a feedback architecture with greater complexity than that of the best possible
non-adaptive compensator. The word “robust” refers to the desire that the adaptive
control system remains stable and also meets the posed performance specifications for
all-possible “legal” parameter values and unmodeled dynamics.
In this chapter, the RMMAC methodology is described in the case of a single
unknown scalar parameter in SISO adaptive problems. The MIMO case, which is
more complex, is described in Chapter 8.
The architecture of the RMMAC system is shown again, for the sake of exposition,
in Figure 5.1.
In order to place our remarks in a proper perspective we repeat the set of basic
engineering questions that naturally arise when we deal with adaptive control:
(1). What do we gain by using adaptive control?
109
110
ξ (t) θ (t)
u (t ) y (t )
r1 (t ) u1 (t )
K1 ( s)
r2 (t ) u2 (t ) u (t )
K 2 ( s)
rN (t ) u N (t )
K N (s )
S1 P1 (t )
S2 P2 (t )
SN PN (t )
(2). How do we fairly predict and compare performance improvements (if any) of
a proposed adaptive design vis-à-vis the “best nonadaptive” one?
(3). How do we design adaptive compensators with guaranteed robust-stability and
robust-performance in the presence of unmodeled dynamics and unmeasurable plant
disturbances and sensor noises?
(4). Is the increased complexity of an adaptive compensator justified by the
performance improvement? What should be the level of complexity for performance
guarantees?
It is important to stress at this point that the vast majority of approaches to
adaptive control deal with the case of constant uncertain real parameters. However,
from an engineering perspective, the true value of an adaptive system can only be
judged by its performance when the uncertain real parameters change “slowly with
time”, within predefined limits.
Thus, one designs and tests an adaptive system for constant real parameters, using
whatever theoretical approaches and/or methodologies are available, but it should be
also tested for time-varying parameters as well.
111
We stress that our design philosophy hinges upon explicit and quantitative speci-
fications for the desired performance (superior disturbance-rejection) of the adaptive
system. The nature and complexity, as well as the required detailed designs, of the
appropriate RMMAC should ideally follow from the posed adaptive performance spec-
ifications. This distinguishes our approach from other adaptive schemes, discussed in
Chapter 2, which focused either upon (local) stability concepts and ad-hoc subjective
performance evaluations, e.g. step-responses and the like.
The remainder of this chapter is as follows. In Section 5.2 we discuss the designs
of the robust dynamic compensators, such as the GNARC, FNARC, and LNARCs,
and how to determine the number N of models in the multiple-model architectures,
as well as the collection of the LNARC compensators. In Section 5.3 we focus upon
the “identification” subsystem of the RMMAC, i.e. designing the Kalman filters.
Section 5.4 provides full details about the proposed RMMAC design methodology.
A technical discussion about the effect of the lower limits on probabilities and stability
of the closed-loop RMMAC are discussed in Section 5.5. In Section 5.6 we introduce
a variant of the RMMAC architecture which “switches in-and-out” the LNARCs. We
call it the RMMAC/S architecture which is more similar to the SMMAC architectures
described in Chapter 2. In Section 5.7 we describe yet another version of the RMMAC,
which we call the RMMAC/XI architecture. The RMMAC/XI designs overcome
potential poor RMMAC disturbance-rejection responses when the plant disturbance
intensity is itself uncertain. Conclusions are summarized in Section 5.8.
(e) Unmeasurable sensor noises, perhaps with information on their power spectral
densities.
113
where x(t) is the state vector, u(t) the control vector, d(t) the plant-disturbance
vector, y(t) the (noisy) measurement vector, n(t) the sensor noise and z(t) the per-
formance (output or error) vector, i.e. the vector for which we wish to minimize the
effects of the disturbance d(t) and noise n(t), i.e. have superior disturbance-rejection.
In this chapter, the system matrices depend upon a real-valued scalar parameter p,
where p is constrained to be in an interval, p ∈ Ω; this is a required µ-synthesis con-
straint. Thus, for each independent real uncertain parameter we must have a lower-
and upper-bound (real-valued).
From a performance point of view, in order to achieve superior disturbance-
rejection, the designer specifies a frequency weight on z(t). Typically, to achieve
superior disturbance-rejection in the low-frequency region, the designer specifies, in
the µ-synthesis methodology, a frequency-weight, say of the form
µ ¶m
α
z̄(s) = Ap · I · z(s) ; m positive integer (5.2)
s+α
To complete the robust design synthesis the designer must provide frequency-
dependent bounds for all (structured or unstructured) unmodeled dynamics, and
frequency weights for the control, disturbance and sensor noise vectors. The control
and sensor noise weights, together with the bound(s) on the unmodeled dynamics,
safeguard against very high-frequency feedback designs.
P = {p : p ∈ Ω ⊂ Rm } (5.3)
Using the ideas of Chapter 4, to compute the GNARC, one fixes the performance
parameter Ap in eq. (5.2) to some initial value and exercises the Matlab software
which, after a sequence of D,G-K iterations, determines a compensator K(s) and
generates an upper-bound µub (ω). K(s) is called the best GNARC if it is associated
with the maximum possible value of performance parameter Ap .
The goal of µ synthesis is to minimize over all stabilizing compensators K(s), the
peak value of µ∆ (·) of the closed-loop transfer function M (P, K). More formally,
If
µub (ω) < 1, ∀ω (5.5)
then the resulting feedback design is guaranteed to be stable for all “legal” unmod-
eled dynamics and the entire parameter uncertainty p ∈ Ω. Moreover, in addition to
stability-robustness, we are guaranteed that we meet or exceed the posed performance
requirements. It should be evident, that in order to find the “best” GNARC we must
maximize the performance-parameter Ap until the µ-upper bound is just below unity,
say, 0.995 ≤ µub (ω) ≤ 1, ∀ω.
The GNARC is a single dynamic compensator that can be used by the designer
to fully understand what is the best possible robust performance in the absence
of adaptation. Since the feedback system is LTI a whole variety of performance
evaluations are possible, using representative values of the uncertain real-parameter
p ∈ Ω for the plant, as summarized in Table 5.1.
Full understanding of the non-adaptive GNARC and its properties are essential
prior to making a decision on using some sort of multiple-model adaptive control.
Moreover, the performance of the GNARC closed-loop system can change drastically if
one adds more measurements and/or controls. Thus, an unacceptable performance for
a SISO GNARC non-adaptive system may, indeed, become acceptable if more controls
and/or sensors are introduced and a MIMO GNARC analyzed, thereby eliminating
the need for complex adaptive control. See Section 6.5, for example.
Finally, as we shall explain in Chapter 10, the GNARC can be used in the design of
a “safety net”, to be used when one “turns-off the adaptive system” by monitoring the
116
(1) Magnitude (or singular-value) Bode plot of the closed-loop transfer function from
the plant disturbance, d, to output, z, which measures the quality of disturbance-
rejection vs frequency. Assuming that the plant has no integrators, the magnitude of
this transfer function, in the low-frequency region, will be approximately 1/Ap . This
is why we must maximize the performance parameter Ap for superior disturbance-
rejection.
(2) Magnitude (or singular-value) Bode plot of the closed-loop transfer function from
the sensor noise, N, to output, z, which measures the quality of insensitivity to sensor
noise vs frequency.
(3) Magnitude (or singular-value) Bode plot of the closed-loop transfer function from
the plant disturbance, d, to control, u, which measures the impact of the plant-
disturbance on the control vs frequency.
(4) Magnitude (or singular-value) Bode plot of the closed-loop transfer function from
the sensor noise, N, to control, u, which measures the impact of the sensor noise on the
control vs frequency.
(5) Root-mean-square (RMS) tables assuming that the plant-disturbance, d, and the
sensor noise, N, are stationary stochastic processes or white noises. Such RMS values
are readily evaluated via the solution of Lyapunov equations. Individual or combined
RMS tables for the output, z, and the control, u, as a function of the plant-disturbance,
d, and the sensor noise, n, can be computed.
If the GNARC analysis discussed above indicates the need for adaptive control, we
have a lower-bound upon robust performance. The “fixed non-adaptive robust com-
pensators (FNARC)” provide the means for quantifying an upper-bound on robust
performance. “Ideally”, the FNARC analysis assumes an infinite number of models,
N → ∞, in any multiple-model adaptive scheme. Thus, we understand what is the
117
Recall that we had initially posed the following question: do we need adaptive control?
The GNARC and FNARC results provide the designer with the tool to answer this
question.
In Figure 5.2 we visualize a hypothetical plot of the outcome of the GNARC and
FNARC designs be plotting the (maximized) performance parameter Ap , see eq. (5.2),
as a function of a scalar uncertain real parameter, p, pL ≤ p ≤ pU . We denote the
(constant maximized) value associated with the GNARC design by AG
p . We denote the
parameter p ∈ [pL , pU ] upon performance. The FNARC process indicates that if the
parameter p were known exactly, then the (low-frequency) disturbance-rejection is
approximately 1/AFp (p). The GNARC process indicates that if the parameter p were
118
Ap
ApF ( p )
re
te FNARC design
am
ra
p
ec
na AGp
m
ro
fr GNARC design
e
P
pL pU
p : Scalar parameter
The shapes of the curves in Figure 5.2 provide additional valuable information.
In the hypothetical case of Figure 5.2, we should expect the benefit of using adaptive
control to be greatest if the parameter was near its upper-bound, i.e. p ≈ pU . The
benefits decrease if the unknown parameter is closer to its lower-bound, i.e. p ≈ pL .
Indeed it may well happen that AFp (pL ) ∼
= AGp . This can occur, for example, if the
If p is a vector, the computations and visualization are more complex; see Chap-
ter 8.
119
A “brute-force” approach is to decide on the number of models, say N =4, and their
parameter variation as illustrated in the hypothetical visualization of Figure 5.3. Es-
sentially, in this approach, the designer fixes the “adaptive complexity”, quantified by
N, of the multiple-model system. Each model, denoted by M#k (k=1,2,3,4) requires
definition of its “local” lower-bound, pkL , and upper-bound, pkU , k = 1, 2, 3, 4, i.e.
Model#1 (M #1) : pL = p1L ≤ p ≤ p1U
Model#2 (M #2) : p1U = p2L ≤ p ≤ p2U
(5.6)
Model#3 (M #3) : p2U = p3L ≤ p ≤ p3U
Model#4 (M #4) : p3U = p4L ≤ p ≤ p4U = pU
The model upper- and lower-bounds must be found by trial-and-error.
After the models are selected, all frequency-dependent bounds and weights are
fixed as in the GNARC and FNARC designs. Next, for each model the performance
parameter Ap – see eq. (5.2) – denoted now by Akp ; k = 1, 2, 3, 4, is maximized using
mixed-µ software, in an iterative mode, for the smaller parameter uncertainty subset
associated with each model. Thus, for each model, we are again attempting to attain
as large a disturbance-rejection as possible. Moreover, at the end of this iterative
optimization process, we also obtain what we call the “local non-adaptive robust
compensator (LNARC)” which we denote by Kj (s), j = 1, 2, 3, 4 (See Figure 5.1).
The values of the Akp ; k = 1, 2, 3, 4, plotted in Figure 5.3 can be used as a “guide”
for adjusting the boundaries of each model or changing the number of models. Each
120
Ap
FNARC
re A4p
te
A3p
am
ra
p
Ap2
ec
na A1p
m
ro
fr
e
P GNARC
M #1 M #2 M #3 M #4 p
pL = p11 p12 = p21 p22 = p31 p32 = p41 p44 = pU
of the resulting LNARC designs can be evaluated in more detail by following the
suggestions in Table 5.1.
The RMMAC numerical results presented in [23, 90] followed such an ad-hoc ap-
proach for determining the models and the associated LNARCs.
The numerical simulations for the RMMAC design in Chapter 6 and 7 utilize a much
more systematic approach to the determination of the number of models and their
numerical specification. This method fully exploits the information provided by the
FNARC in Figure 5.2. We illustrate the idea for one scalar uncertain parameter.
We have repeatedly emphasized that the number of models required for any adap-
tive multiple-model design should be the natural outcome of the posed performance
design specifications. In the so-called % FNARC approach the designer specifies that
the performance parameter, Ap , should be equal or greater than X% of the best
possible performance as defined by the FNARC.
The basic idea is illustrated in Figure 5.4, starting from the GNARC and FNARC
121
FNARC
ApF ( p )
Ap X% of FNARC
re
1
te 2
m
ar
ap
ec
an
GNARC
m
ro
rfe
P
Ω2 Ω1
M #2 M #1
p
pL . . . β* α* pU
curves of Figure 5.2. Using the designer-specified value of X%, we construct the
X% FNARC, shown in Figure 5.4, and we proceed as follows. Since the FNARC
is maximum at p = pU , we slowly increase, starting from the upper limit, the size
of the parameter uncertainty set Ωα = {p : α ≤ p ≤ pU }. For each value of α, we
use the mixed-µ software to design the best robust compensator by maximizing the
performance parameter Ap denoted by Aαpmax . As long as Aαpmax > (X%) · AFp (α),
where AFp (α) is the parameter value of the FNARC at p = α, then we decrease α
∗
until at α = α∗ we have Aαp max = (X%) · AFp (α∗ ). The outcome of this process defines
the dashed curve labeled Γ1 in Figure 5.4. The point α∗ is at the intersection of the
Γ1 curve with the X% FNARC curve. This defines Model#1 (M#1) with uncertainty
set Ω1 = {p : α∗ ≤ p ≤ pU }; we also stress that the µ-software also determines the
LNARC#1 compensator denoted by K1 (s). The process is repeated from the (right)
boundary of M#1, α∗ . Starting at p = α∗ , we define the set Ωβ = {p : β ≤ p ≤ α∗ }.
For each value of β, we use the mixed-µ software to design the best robust compensator
by maximizing the performance parameter Ap denoted by Aβp max . As long as Aβp max >
(X%) · AFp (β), where AFp (β) is the parameter value of the FNARC at p = β, then
∗
β is decreased until at β = β ∗ we have Aβp max = (X%) · AFp (β ∗ ). The outcome
of this process defines the dashed curve labeled Γ2 in Figure 5.4. The point β ∗
is at the intersection of the Γ2 curve with the X% FNARC curve. This defines
122
Li
L2 = L( p, p*2 )
L1 = L( p, p1* )
M
PB
p1* pA p2* p
Fundamental convergence result: In [44] this result was generalized and proved
for an arbitrary number of KFs designed for the nominal values p∗1 , p∗2 , . . . , p∗N . If the
BPM satisfies the inequality
then, under some additional stationarity and ergodicity assumptions – see Chapter 3
– the posterior probabilities converge almost surely to the correct model, i.e.
To compensate for the unknown parameter within each KF we may need to use
some “fake plant noise,” so that the KF pays more attention to the information
contained in the residual signal.
The outcome of the compensator design of Section 3 yielded the number of models,
N, and their boundaries, in terms of the real subsets Ω1 , Ω2 , . . . , ΩN which define the
125
Li
L3
L2
M
PB
L1
Figure 5.6: Optimizing the KF nominal design points using the BPMs.
models M #1, M #2, . . . , M #N. The question now is: how do we select the nominal
values p∗1 , p∗2 , . . . , p∗N to design the KFs?
The idea, illustrated for three models in Figure 5.6 for a scalar parameter, is to
use an iterative algorithm to calculate the nominal KF values p∗1 , p∗2 , p∗3 , so that the
BPMs are equal at the boundary of adjacent subintervals (models) Ωs .
In this manner, the fundamental probability convergence result will guarantee
that
p ∈ Ωj ⇒ Pj (t) → 1 a.s.
This is the method we use in the numerical RMMAC simulations of Chapters 6 and
7. This method becomes more complicated when we have two, or more, uncertain
parameters – see Chapter 8.
ξ (t) θ (t)
u (t ) y (t )
r1 (t ) u1 (t )
K1 ( s)
r2 (t ) u2 (t ) u (t )
K 2 ( s)
rN (t ) u N (t )
K N (s )
S1 P1 (t )
S2 P2 (t )
SN PN (t )
(1). Independent of the size of uncertainty for the plant real parameter, the
plant always contains unmodeled dynamics whose size can be bounded a priori only
in the frequency domain. Therefore, the adaptive design must explicitly reflect these
frequency-domain bounds upon the unmodeled dynamics. The presence of unmodeled
dynamics immediately brings into sharp focus the fact that we must use the state-of-
the-art nonadaptive robust control synthesis, i.e. mixed-µ synthesis [1–3, 47, 107, 110,
114]; and associated Matlab software [3, 114].
(2). The plant is subject to unmeasurable plant disturbances whose impact upon
the chosen performance variables (error signals) must be minimized, i.e. we must have
superior “disturbance-rejection”. The modern trend is to use frequency-depended
weights to emphasis and define superior disturbance-rejection performance. This
design objective can also be accommodated by the mixed-µ design methodology.
(3). The plant measurements are not perfect; thus sensor measurements are cor-
rupted by unmeasurable sensor noise. The performance variables must be “insensi-
tive”, to the degree possible, to such sensor noise.
(4). Performance requirements must be explicitly defined, up to constants whose
127
values can be optimized for superior performance. In the mixed-µ design methodology
these “disturbance-rejection” performance requirements are explicitly quantified by
frequency-domain weights typically involving the selected “error signals” and the
control variables.
(5). Given the information in (1) to (4), we can design the best “global non-
adaptive robust compensator (GNARC)” for the entire (large) uncertain real-parameter
set and taking into account both the unmodeled dynamics and the performance re-
quirements. This non-adaptive feedback design must be optimized so as to yield the
best possible performance, i.e. superior disturbance-rejection with reasonable control
effort. The GNARC then provides a yardstick (lower-bound) for performance, so that
any performance improvements by more complex adaptive designs can be quantified.
It is self-evident that the GNARC must be designed using the mixed-µ synthesis
methodology.
(6). An upper-bound for adaptive performance can be obtained by optimizing
the performance under the assumption that the real-parameter value is known ex-
actly, but still reflecting the presence of complex-valued unmodeled dynamics and
frequency-dependent performance requirements. This implies that we compute for a
large number of grid points in the original parameter uncertainty set a “fixed non-
adaptive robust compensator (FNARC)” which defines the best possible performance
for each parameter value. The FNARC design is carried out using the complex-µ syn-
thesis methodology, since we still must take into account the unmodeled dynamics
and frequency-dependent performance specifications. We use the same quantitative
performance requirements as in part (5) above. The set of the FNARCs corresponds
to having an infinite number of models in the multiple-model implementation.
The difference between the lower-bound on performance from the GNARC in part
(5) and the FNARC upper-bound from part (6) provides a valuable quantitative de-
cision aid to the designer on what performance improvements are possible by some
multiple-model adaptive control method. The designer must then make a quantifiable
choice on the degree of performance improvement that he/she desires from the adap-
tive system. As we already discussed, this approach will then define the number of
models (parameter subsets) required, N, and their numerical specification (boundary
128
y(t) u1(t)
K1
D x
q p P1 (t )
u2(t) u(t)
K2 x +
w z
P P2 (t )
u y
uN(t)
KN x
K M(P,K) PN (t )
From PPE
5.4.1 Discussion
The RMMAC has the following common characteristics.
(a). One must design a set of N dynamic compensators (LNARCs).
(b). One must design a set of N Kalman filters.
(c). One must implement an identification process by which the actual (global)
adaptive control is generated.
Clearly, the complexity of the adaptive system will depend on the number, N,
of models that are required to implement. Ideally, N should be as small as possible.
However, it should be intuitively obvious that if N is too small the performance of the
adaptive system may not be very good. On the other hand, if N is very large, one may
reach the point of diminishing returns as far as adaptive performance improvement
is concerned. It follows that we need a systematic procedure by which, starting
from a compact parameter set p ∈ Ω, to define a finite set, N, of discrete values
(models), ΩD = {p1 , p2 , . . . , pN } ⊂ Ω, that are used subsequently in designing the
KFs in the RMMAC, as well as the compensators. The following section summarizes
our suggested methodology which hinges upon the recent developments in robust
feedback control synthesis using the so-called mixed-µ methodology and associated
software.
131
– For each value of the real parameter, the performance parameter, Ap , is again
maximized until µ upper-bound ∼ = 0.995.
Repeat Step 3 for a grid of the real parameter that cover the initial parameter
uncertainty set Ω
Plot the (maximized) Ap from the GNARC and the FNARC designs vs p as in
Figure 5.2.
• STEP 6. [Selection of X% ]
Now the minimum number, N, of required models and their uncertainty range
can be obtained as illustrated in Figure 5.4 (see also examples in Chapters 6
and 7.)
– We use the same frequency weights as in the GNARC, so that the performance
comparisons are “fair”.
– Each LNARC will guarantee a “local” performance that exceeds that of the
GNARC.
It is assumed that the parameter varies over a continuous region and a finite
number N (obtained in Step 7) of constant-gain KF are available for the esti-
mation.
– The KFs are NOT used for state-estimation, only for “identification” via the
PPE.
– From Step 7 we know the (smaller) parameter subinterval associated with each
model indexed by k; k = 1, 2, . . . , N . To design each KF we must determine a
“nominal” parameter value, p∗k .
– The desired p∗k ’s are obtained by matching values of the BPM between adjacent
parameter subintervals; see Figure 5.6 and examples in Chapters 6 and 7.
Assuming that the RMMAC probabilities converge, we can evaluate for each
LNARC the detailed expected performance, vis-a-vis that of the GNARC, using:
However, the uncertain parameter of the true model can be time-varying, and
thus one would like to require the posterior probabilities Pi to be nonzero (but small)
for all time so that the RMMAC algorithm can respond quickly to such parameter
changes. Thus, one almost always applies an additional constraint on the probability
such as
Pi (t) ≥ Plim ; ∀i = 1, . . . , N (5.13)
This has been done in all examples throughout this thesis, in Chapters 6, 7, and
9, with a value of Plim = 10−14 . The effect of such a small limit, does not degrade
the performance of the RMMAC as it will be evidenced by numerous simulations.
However, selecting a large Plim may affect the performance of the RMMAC. Such a
situation must be avoided as clarified in more detail in the next section.
To place an artificial lower limit on the probability of any model can be accom-
plished by the inclusion of a Simulink statement such as
if Pi (t) ¹ Plim ,
Pi (t) = Plim ; (5.14)
end
This prevents the probability of any model to become exactly zero due to round-
off errors. This is detrimental not only in cases in which the true model is slowly
changing but also in cases in which an oscillatory behavior occurs (e.g. in cases
that the theoretical assumptions are harshly violated), because roundoff may occur
135
before the probability switches. It should be recognized that although the lower limit
is somewhat artificial, it is required in some form whenever a computer with finite
precision is used for control calculations.
The existence of a lower limit for adaptive case is an essential issue and it may
require the modification of the analysis of the behavior of the RMMAC.
ẋp = Ap xp + Bp u + Lp d
(5.15)
y = C p xp + θ
ẋ1 = A1 x1 + B1 y
(5.16)
u1 = C1 x1
ẋ2 = A2 x2 + B2 y
(5.17)
u2 = C2 x2
Now, consider the RMMAC system with above unknown plant and the two com-
pensators as shown in Figure 5.9.
3
The fact that some LNARCs can be individually destabilizing is commonplace in RMMAC
designs, as we shall see in Chapters 6, 7, and 9.
136
In Figure 5.9, α1 and α are in fact the steady-state values of the posterior probabil-
ities that weigh the local control signals of LNARC#1 (K1 (s)), u1 (t), and LNARC#2
(K2 (s)), u2 (t). If the theoretical assumptions are satisfied, α1 ≈ 1 and α2 ≈ 0. How-
ever, due to the effect of the probability lower limit (discussed in Section 5.5.1) we
shall show them as two constant (non-zero) gains as follows.
and
u2 = C2 (SI − A2 )−1 B2 y (5.19)
| {z }
F2 (s)
= P u + Md + θ (5.20)
u = α1 u1 + α1 u2
= (α1 F1 + α2 F2 )y
| {z }
F
= F [P u + M d + θ] (5.21)
137
u = (I − F P )−1 F (M d + θ) (5.22)
y = P (I − F P )−1 F (M d + θ) + M d + θ
£ ¤ (5.23)
= I + P (I − F P )−1 F (M d + θ)
It turns out that the closed-loop system, with two stochastic inputs, is stable if
and only if
£ ¤−1
I + P (I − F P )−1 F 6= 0
or
£ ¤
det I + P (I − F P )−1 F 6= 0
Now, assume that the lower bounds of the probabilities are very small, i.e. Plim ≈
². Then we have
F = α1 F1 + α2 F2 ≈ F1
and
I − F P = I − (α1 F1 + α2 F2 )P
= I − α1 F1 P − α2 F2 P
≈ I − F1 P
and hence
£ ¤
y ' I + P (I − F1 P )−1 F1 (M d + θ)
that is the same output of the closed-loop system when only the stabilizing compen-
sator K1 (s) is placed in the feedback with the unknown plant in a single separate
loop.
It turns out that the poles of the whole closed-loop system, even if the destabilizing
compensator K2 is in the feedback, are “almost” at the same location of the close-loop
poles when only the stabilizing compensator K1 is the only one used. It means that
the global control signal is almost entirely affected only by the stabilizing compensator
138
K1 (s). Even if the destabilizing compensator K2 is in the feedback loop, it will not
affect the stability, as long as α2 ≈ 0. Thus, the lower limit on probabilities should
be selected sufficiently small such that the unstable compensators would not able to
change the location of the poles of the closed-loop system.
Roughly speaking, the compensator K2 (s) destabilizes the plant in isolation. How-
ever, in the RMMAC context, K2 (s) faces the feedback combination of P and the
stabilizing K1 (s) and it contains an infinitesimal gain α2 ≈ 0 as well, as illustrated in
Figure 5.9.
One can easily show that the signals of the destabilizing compensator are also
bounded if a sufficiently small lower bound Plim is selected. It is also easy to show
that this is also the case where the RMMAC has three or more LNARCs.
ξ (t) θ (t)
u (t ) y (t )
r1 (t ) ⇓ u1 (t )
r2 (t ) u 2 (t ) u (t )
rN (t) u N (t )
S1 P1 (t )
S2 P2 (t ) σ k ( Pk )
SN PN (t )
to reflect different ranges of the plant white noise intensity Ξ. Let us assume that the
number, N , of models and their size has been determined by the required adaptive
performance specifications, as we described in this chapter, and that we have already
calculated the N LNARCs denoted by Kk (s) ; k = 1, 2, . . . , N . Let us further suppose,
for the sake of exposition, that the plant noise intensity Ξ is scalar and is bounded
by
ΞL ≤ Ξ ≤ ΞH
and that, for the sake of simplicity, we decide to select two intermediate values, Ξ1
and Ξ2 , such that
ΞL < Ξ1 < Ξ2 < ΞH
Then we design two sets of linear KFs, one set indexed by k = 1, 2, . . . , N using the
noise intensity Ξ1 and the second set indexed by k = N + 1, N + 2, . . . , 2N using the
noise intensity Ξ2 . It is important to stress that the RMMAC/XI structure is valid for
the MIMO case as well (see also Section 9.7.6). The nominal points for designing each
KF, as determined by the BPM method of Sections 5.3 and 8.3, will be different (and
must be recomputed) for each value of Ξ1 and Ξ2 . Essentially, we have doubled the
number of hypotheses in the associated MMAE; but we can still apply the MMAE-
based identification methodology. This allows us to define the RMMAC architecture,
which we shall refer to as RMMAC/XI, shown in Figure 5.12.
Note that each KF will now generate a different on-line residual, rk (t) ; k =
1, 2, . . . , 2N . Also, we can still pre-compute the 2N residual covariance matrices Sk .
These are introduced to the PPE of Figure 5.12 which will generate 2N posterior
probabilities, Pk (t) ; k = 1, 2, . . . , 2N , one for each of the 2N hypotheses. Note,
however, that the bank, and number N, of the N LNARCs is unchanged, since their
design does not depend on Ξ. Thus, it is important to note that each LNARC control
signal uk (t) ; k = 1, 2, . . . , N , is multiplied by the two different associated posterior
probabilities, Pk (t) and Pk+N (t), to generate the overall adaptive control u(t).
It is obvious how to generalize this idea for more than two values of Ξ. It is an
open research question on how to select “intelligently” the values of Ξ, other than by
brute-force and trial-and-error.
142
ξ (t); Ξ ∈ {Ξ1 , Ξ 2 }
θ(t )
u (t ) y (t )
r1 (t ) u1 (t )
K1 ( s )
Designed r2 (t )
with
Ξ=Ξ
1 u2 (t ) u (t )
K 2 ( s)
rN (t )
rN +1 (t )
u N (t )
K N ( s)
Designed rN + 2 (t )
with
Ξ=Ξ
2
r2 N (t )
S1 P1 (t )
S2 PN (t )
PN +1 (t )
S2 N
P2 N (t )
6.1 Introduction
145
146
e − sτ
Figure 6.1: The two-cart system. The spring stiffness, k1 , is uncertain, and there is
an unmodeled time-delay in the control channel.
The system in Figure 6.1 includes a random colored disturbance force d(t), acting
147
on mass m2 , and sensor noise on the only measurement available, which is that of the
position of mass m2 . The control force u(t) acts upon the mass m1 .
The disturbance force d(t) is modeled by a stationary stochastic process generated
by driving a low-pass filter, Wd (s), with continuous-time white noise ξ(t), with zero
mean and unit intensity, as follows:
with
α
Wd (s) = (6.2)
s+α
where the frequency range ω ≤ α is where the disturbance has most of its power.
The overall state-space representation, including the disturbance dynamics via the
state variable x5 (t), is:
and
0 0 1 0 0
0 0 0 1 0
A = − m k 1 k 1
− b1 b 1
0
1 m1 m1 m1
k1 (k +k ) (b +b )
− 1 2 b1
− 1 2 1
m2 m 2 m 2 m 2 m 2
0 0 0 0 −α
h i h i
B T = 0 0 m1 0 0 ; C = 0 1 0 0 0
1
h i
T
L = 0 0 0 0 α
The upper- and lower-bound for the uncertain spring constant, k1 , are
All feedback loops utilize a single measurement y(t), the position of mass m2 , that
includes additive white sensor noise θ(t), defined by
The desired disturbance-rejection requires that the effects of d(t) (primarily) and
also θ(t) be minimized so that z(t) ≈ 0.
Remark: The control problem is hard even if the spring constant k1 is known. This
is a non-collocated actuator problem because the control is not applied directly to the
mass m2 whose position is to be regulated. Clearly, the control problem becomes
even harder in our adaptive design, because the control u(t) is applied through the
uncertain spring, so we are not sure how much force is exerted through the uncertain
spring k1 on the mass m2 .
The frequency responses of the open-loop of the MSD system, from d(t) to z(t)
and also from control u(t) to z(t), are shown in Figure 6.2 for different values of k1 ; see
eq. (6.5). Note that the uncertainty in k1 is most noticeable in the higher frequency
mode.
In addition to the uncertain spring stiffness, we assume that there is, in the control
channel, an unmodeled time-delay τ whose maximum possible value is 0.05 secs, i.e.
2
10
k =1.75
1
Mag.
0
10
k1=0.25
−2
10
−2 −1 0 1
10 10 10 10
Frequency (rad/sec)
(a) From disturbance d(t) to output z(t)
2
10
0 k1=1.75
10
Mag.
k =0.25
−2 1
10
−2 −1 0 1
10 10 10 10
Frequency (rad/sec)
(b) From control u(t) to output z(t)
Figure 6.2: Bode plot of the open-loop MSD system for different values of the spring
constant k1 , eq. (6.5).
1
10
Unmodeled time−delay
Frequency bound
0
10
Magnitude
−1
10
−2
10
−3
10
−1 0 1 2 3
10 10 10 10 10
Frequency (rad/cec)
Figure 6.3: Frequency weight of unmodeled time-delay dynamic used in the GNARC
and each LNARC designs and its frequency bound, for τ = 0.05 secs.
α
Wp (s) = Ap (6.11)
s+α
Hence, from Section 6.2, the 1st order disturbance modeling filter is
0.1
Wd (s) = (6.13)
s + 0.1
As is customary in H2 or H∞ designs, we also use the following frequency-domain
weights on the control and measurement noise.
10(s + 10)
Control weight: Wu (s) =
s + 103 (6.14)
−3
Measurement noise weight: Wn = 10 (constant)
These, together with the unmodeled dynamics weight (6.9), limit the bandwidth
of the closed-loop system by penalizing large high-frequency control signals. The
weights of eqs. (6.9) and (6.14) are incorporated in the definition of the nominal
generalized plant, together with eq. (6.3) with a nominal k1 = 1.0. The weights of
eqs. (6.13), (6.9), and (6.14) will not change in any of the subsequent designs.
Figure 6.4(a) shows the low-pass 1st order disturbance modeling filter. Note
that from eqs. (6.13) and (6.11), Figure 6.4(a) is also the plot of the normalized
performance weight Wp (s) where Ap is assumed to be unity. Figure 6.4(b) shows the
frequency response of the selected control weight. Its general behaviour is due to the
fact that the performance output of the disturbance-rejection problem was selected
for those frequencies in which plant disturbances have most of their power.
By comparing Figures 6.2(a) and Figures 6.4(a) we see that in the LFD the noise
power is below the first resonant mode of the MSD.
Following the step-by-step algorithm in Chapter 5, the GNARC and the LNARCs
are designed next.
1
0
10 10
−1
10
Mag.
Mag.
0
10
−2
10
−3 −1
10 10
−2 −1 0 1 2 0 2 4
10 10 10 10 10 10 10 10
Freq (rad/sec) Freq (rad/sec)
(a) The 1st order disturbance modeling (b) Control weight Wu (s), a high-pass fil-
low-pass filter, Wd (s) ter that penalizes the system for using
large controls at high frequencies
for the entire large parameter uncertainty of eq. (6.5). As explained in Section 5.2.1,
the GNARC system defines what one can best expect in the absence of adaptation.
To carry out the mixed-µ synthesis, the parameter uncertainty of eq. (6.5) is
represented by
In order to design the “best possible” non-adaptive feedback system the following
performance weight upon the output z(t) is used.
0.1
Wp (s) = Ap (6.16)
s + 0.1
which reflects our specification for good disturbance-rejection for the frequency
range ω ≤ 0.1 rad/sec where the disturbance d(t) has most of its power. We again
stress that the performance weight Wp (s) penalizes output error in the same frequency
range as the disturbance dynamics Wd (s) while the gain parameter Ap in Wp (s)
specifies our desired level of disturbance-rejection. The larger Ap , the greater the
penalty on the effect of the disturbances on the position of mass m2 . As described
in Chapter 5, for superior disturbance-rejection, Ap should be as large as possible;
how large it can be is limited by the required guarantees on robust-stability and
-performance inherent in the mixed-µ synthesis methodology.
Figure 6.5 shows the MSD plant with the weights as required by mixed-µ synthesis.
One can note that there are two frequency-weighted “errors” z̃(t) and ũ(t). This figure
153
ξ (t ) θ (t )
Wd(s) Wn
Plant model set
Wp(s)
∆un(s) d (t ) δ k1
Wun(s) α
Ap
z (t )
s +α
+ G(s) z (t )
+
u (t )
Wu(s)
u (t ) y (t )
K(s)
Figure 6.5: MSD system with weights for the mixed-µ synthesis.
is in fact a block diagram of the uncertain closed-loop MSD system illustrating the
disturbance-rejection performance objective namely the closed-loop transfer function
from ξ(t) → z(t), or d(t) → z(t).
The “position error” z̃ is our main performance variable for evaluating the quality
of the disturbance-rejection. Since
µ ¶
0.1
z̃(s) = Wp (s) z(s) = Ap z(s) (6.17)
s + 0.1
we communicate to the µ-design that position errors are most important below
the “corner frequency” 0.1 rad/sec. The larger the performance parameter Ap , the
more one cares about position errors at all frequencies.
The “control error” ũ(t) is defined by
µ ¶
s + 10
ũ(s) = Wu (s) u(s) = 10 u(s) (6.18)
s + 103
Thus, Wu (s) is a high-pass filter that penalizes the system for using large controls
at high frequencies.
Using the mixed-µ software the performance parameter Ap in eq. (6.16) is in-
creased, as much as possible, until the upper-bound on the mixed-µ, µub (ω), just
satisfies the inequality
µub (ω) ≤ 1, ∀ω (6.19)
154
AG
p = 50.75 (with µub ≈ 0.995) (6.20)
which leads to the GNARC design, the “best” LTI non-adaptive compensator K(s)
that guarantees stability- and performance-robustness for the entire parameter in-
terval as in eq. (6.5). As explained in Section 5.2.1, the performance characteristics
of the GNARC are to be used as a comparison-basis for evaluating performance im-
provement (if any) of the proposed RMMAC design. See Section 5.2.1 for detailed
discussions.
3
10
GNARC FNARC
p
A
2
10
1
10
0.25 0.75 1.25 1.75
k1
Figure 6.6: Best GNARC and FNARC performance gains for spring-stiffness uncer-
tainty as in eq. (6.5). Note the logarithmic scale.
155
Figure 6.7 shows how four covering models are defined to be adequate for con-
structing the RMMAC using the performance level of 70%.
FNARC
1000
800
600
70% of FNARC
p
A
400
200
Ω4 Ω3 Ω2 Ω1
GNARC
0
0.25 0.5 0.75 1 1.25 1.5 1.75
M#4 M#3 M#2 Model #1(M #1) k1
Figure 6.7: Best local performance gains using mixed-µ using 70% of the best FNARC
performance designed for local uncertainty bounds.
As illustrated in Figure 6.7 for the specified performance level of 70%, the large
parameter uncertainty interval of eq. (6.5) is subdivided into four subintervals as
summarized in Table 6.1.
Explanation. We now provide some more details on how the four models of Fig-
ure 6.7 are obtained. Starting at the North-East corner labeled F1 , k1 = 1.75, (because
the FNARC is maximum there) we slowly open-up the uncertain interval denoted by
Ωb = {k1 : b ≤ k1 ≤ 1.75}. For each interval, we iterate, using the mixed-µ software,
to calculate the maximum performance parameter Abp which results in the dashed
curve, labeled Γ1 in Figure 6.7. When the curve Γ1 intersects the 70% FNARC
curve we stop. In this example, this occurs at k1 = 1.02 which yields the subset
Ω1 = [1.02, 1.75], i.e. 1.02 ≤ k1 ≤ 1.75. This yields Model#1 (or M#1), and also
the compensator K1 (s) (or LNARC#1). The left boundary of M#1 defines the point
labeled F2 on the FNARC. Starting from this point we again slowly open up the
interval denoted by Ωc = {k1 : c ≤ k1 ≤ 1.02}. Once more, we iterate using the
mixed-µ software to calculate the maximum performance parameter Acp which results
156
in the dashed curve, labeled Γ2 in Figure 6.7. When the curve Γ2 intersects the 70%
FNARC curve we stop. In this example, this occurs at k1 = 0.64 which yields the
subset Ω2 = [0.64, 1.02], i.e. 0.64 ≤ k1 ≤ 1.02, Model#2, and LNARC#2. Repeating
this process leads to the curves labeled Γ3 and Γ4 in Figure 6.7 and the four models
summarized in Table 6.1.
As a result, the entire initial uncertainty set of eq. (6.5) is covered by four models
requiring four local compensators. Clearly, the reduction in parameter uncertainty
allows larger performance gains for designing the LNARCs, resulting into guaranteed
both stability- and -performance robustness over the subintervals of Table 6.1.
In fact the above model selection procedure using mixed-µ synthesis also generates
four “local” robust compensators, K1 (s), . . . , K4 (s), designed for each subinterval
defined in Table 6.1; these are referred to as LNARCs. In the mixed-µ synthesis, the
weights eq. (6.13), eq. (6.9) and eq. (6.14) were the same as in the GNARC design of
Section 6.3.1. However, for each LNARC design, the performance parameter Ap in
eq. (6.16) is maximized until the mixed-µ upper-bound of eq. (6.19) is achieved, and
their optimized values are shown in the last column of Table 6.1.
Figure 6.8 compares the GNARC design with the four LNARCs, namely K1 (s),
. . . , K4 (s) via their Bode magnitude plots. Note that at low frequencies the LNARCs
generate a loop-gain about 20 times as large compared to the GNARC and this,
naturally, leads to the performance improvements discussed below.
We emphasize that each individual LNARC closed-loop design has guaranteed
performance- and stability-robustness over its associated parameter subinterval of
Table 6.1.
157
4
10
3 LNARCs
10
2
10
Mag.
GNARC
1
10
0
10
−1
10 −2 0 2 4
10 10 Freq. (rad/sec)10 10
Remark: In the design philosophy adopted in this thesis we have stressed that the
adaptive compensator complexity, as measured by the number of models in the CM-
MAC, SMMAC and RMMAC should be the natural by-product of the performance
requirements. We have just demonstrated that if we demand that the performance
equals or exceeds X% = 70% of the FNARC, we require the four models summarized
in Table 6.1.
If one is willing to have somewhat inferior performance and select, say, X% = 50%
then the procedure outlined results in only two models. If we wish to have much
better performance and select, say, X% = 90%, then the outlined procedure yields
nine models. Clearly, as we demand better and better performance we must increase
the RMMAC complexity, and this agrees with engineering intuition.
Testing the RMMAC requires significant computation using multiple Monte Carlo
(MC) runs under different scenarios. It is also important, as explained in Section 5.2,
to use the LTI feedback designs, using the GNARC and LNARCs, to quantify the
potential benefits of using adaptive control in general, and the RMMAC in particular.
From a pragmatic engineering perspective we must have tradeoffs that contrast the
performance improvements (if any) of the very sophisticated RMMAC vis-a-vis the
158
much simpler non-adaptive GNARC design. To the best of our knowledge, such
performance tradeoffs have not been quantified in other adaptive control studies.
Figure 6.9 shows the four-model RMMAC architecture for this example (simply
a specialization of Figure 5.7). Clearly, the RMMAC in Figure 6.9 requires running
four KFs, four dynamic LNARCs, K1 (s), . . . , K4 (s), and computing the four prob-
abilities, P1 (s), . . . , P4 (s) using the posterior probability evaluator (PPE) – a lot of
computations!
ξ (t) θ (t)
u (t ) y (t )
Unknown plant
r1 (t ) u1 (t ) u (t )
KF#1 K1(s)
r2 (t ) u2 (t )
KF#2 K2(s)
r3 (t ) u3 (t )
KF#3 K3(s)
r4 (t ) u4 (t )
KF#4 K4(s)
S1 P1 (t )
Residual S2 Posterior P2 (t ) Posterior
covariances Probability P3 (t ) probabilities
S3
Evaluator
S4 (PPE) P4 (t )
Figure 6.9: The four-model RMMAC architecture for the MSD example (LFD case).
In order to understand how one can easily predict the potential RMMAC perfor-
mance characteristics, assume that one of the posterior probabilities converges to its
nearest probabilistic neighbor (which it does, as we demonstrate in the sequel when
we do not severely violate the theoretical assumptions); it follows that a specific
LNARC is used. After the probability convergence, the RMMAC essentially operates
as an LTI stochastic feedback system!
This allows us to calculate two key transfer functions for disturbance-rejection
159
z(s) z(s)
Disturbance - rejection: Mξz (s) ≡ or Mdz (s) ≡
ξ(s) d(s)
(6.21)
u(s) u(s)
Control - signal: Mξu (s) ≡ or Mdu (s) ≡
ξ(s) d(s)
for different values of the unknown spring stiffness of eq. (6.5), for the GNARC
and for each LNARC design.
Figure 6.10 illustrates the above using the actual spring constants, the nominal val-
ues of KFs1 , quantifying the potential RMMAC improvement in disturbance-rejection
in the frequency domain. Similar plots can be made for other values of the uncertain
spring stiffness. Figure 6.10 predicts that the RMMAC has the potential to signif-
1
icantly improve low-frequency disturbance-rejection on the order of Ap
as expected;
these predictions will be validated in the sequel.
1 1
10 10
0 0
10 10
−1 −1
10 10
GNARC
Mag.
GNARC
Mag.
−2 −2
10 10
LNARC #1 LNARC #2
−3 −3
10 10
−4 −4
10 10
−3 −2 −1 0 1 2 −3 −2 −1 0 1 2
10 10 10 10 10 10 10 10 10 10 10 10
Freq. (rad/sec) Freq. (rad/sec)
(a) GNARC and LNARC#1 for k1 = 1.20 (b) GNARC and LNARC#2 for k1 = 0.75
1 1
10 10
0 0
10 10
−1 −1
10 10
GNARC GNARC
Mag.
Mag.
−2 −2
10 10
LNARC #3 LNARC #4
−3 −3
10 10
−4 −4
10 10
−3 −2 −1 0 1 2 −3 −2 −1 0 1 2
10 10 10 10 10 10 10 10 10 10 10 10
Freq. (rad/sec) Freq. (rad/sec)
(c) GNARC and LNARC#3 for k1 = 0.47 (d) GNARC and LNARC#4 for k1 = 0.30
Figure 6.10: Potential improvement from RMMAC visualized by the Bode plots of
disturbance-rejection transfer function |Mdz (jω)|.
1
These are calculated in the next subsection.
160
0
10
0
10
−1
10
Mag.
Mag.
−1
−2 10
10
GNARC GNARC
−3
LNARC#1 −2
LNARC#4
10 10
−3 −2 −1 0 1 2 −3 −2 −1 0 1 2
10 10 10 10 10 10 10 10 10 10 10 10
Freq. (rad/sec) Freq. (rad/sec)
(a) TF from ξ(t) to u(t) for k1 = 1.20 (b) TF from ξ(t) to u(t) for k1 = 0.30
4 4
10 10
GNARC GNARC
LNARC#1 LNARC#4
2 2
10 10
Mag.
Mag.
0 0
10 10
−2 −2
10 10
0 5 0 5
10 10 10 10
Freq. (rad/sec) Freq. (rad/sec)
(c) TF from θ(t) to u(t) for k1 = 1.20 (d) TF from θ(t) to u(t) for k1 = 0.30
Figure 6.11: Bode plot of transfer functions from ξ(t) or θ(t) to control u(t).
Similar plots, such as control signal characteristics, can be made for all possi-
ble values of the uncertain spring constant. Some of the transfer functions (TFs)
computed from the disturbance ξ(t) and the sensor noise θ(t) to the control u(t) are
shown in Figure 6.11. The areas under these curves show that the LNARCs should
use slightly more RMS of control signals to yield the improved performances which
are shown in Figure 6.10.
Figure 6.12 evaluates the potential performance improvement of using RMMAC,
due to ξ(t) only, by using stochastic metrics, namely by comparing the RMS errors
of the output z and the control u, for different values of k1 .
Assuming that ξ is indeed white noise, these RMS results are readily computed by
solving standard covariance algebraic Lyapunov equations for stochastic LTI systems
[16, 56, 94]. The graphs of Figure 6.12(a) vividly suggest that RMMAC has the
potential of decreasing the output RMS by a factor of 2 to 5 over the GNARC
system. Note that the potential improvement in output RMS requires controls with
higher RMS values, as expected in Figure 6.11.
Similarly, Figure 6.13 evaluates the potential performance improvement of using
161
−2
×10
8
RMS of z(t)
4 GNARC
2 LNARC#3
LNARC#2
LNARC#1
1
GNARC
0
0.25 0.75 k1
1.25 1.75
Figure 6.12: Predicted potential RMS performance of the RMMAC vs GNARC, with
no sensor noise, i.e. θ(t) = 0.
× 10−2
8
GNARC
6 LNARC
RMS of z(t)
0
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
k1
15
14.5
14
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
k1
Figure 6.13: Predicted potential RMS performance of the RMMAC vs GNARC with
both ξ and θ.
162
RMMAC, now in the presence of sensor noise, θ(t), by comparing the RMS errors of
the output z and the control u, for different values of k1 .
Assuming that ξ and θ are indeed white noises, again these RMS results are
readily computed by solving standard covariance algebraic Lyapunov equations for
stochastic LTI systems. Figure 6.13(a), similar to Figure 6.12(a), strongly suggests
that RMMAC has the potential of decreasing the output RMS by a factor of 2 to
5 over the GNARC system. However, in the presence of sensor noise, the potential
improvement in output RMS will require larger control signals than with θ = 0;
compare Figure 6.12(b) and Figure 6.13(b).
Finally, it is important to construct what we call the “stability-mismatch” table
shown in Table 6.2. The following notation is used in Table 6.2:
The interpretation of Table 6.2 answers the question: what happens to the closed-
loop stability if we use the LNARC#j, Kj (s), when the true spring constant is in
subinterval #i? The diagonal entries in this table are always robustly-stable, by con-
struction. Examining the first row in Table 6.2 it is observed that for Model#1, i.e.
for all k1 ∈ [1.20, 1.75], if we use the LNARC#4 we always have instability (U); if we
use the LNARC#2 we have instability for smaller values of k1 , but have stability for
larger values (CU). This is due to the fact that the mixed-µ upper-bound inequal-
ity (6.19) is only a sufficient condition for both robust-stability and -performance,
163
Controller coverage
K1(s)
M#1
K2(s)
M#2
K3(s)
M#3
K4(s)
M#4
0.25 0.55 0.85 1.15 1.45 1.75
k
1
Figure 6.14: The mismatch-instability cases where each local compensator Kj (s),
LNARC#j, is tested to be placed in the feedback while the uncertain parameter
(spring constant k1 ) varies over all possible values in the set, Ω.
×: shows an unstable model/compensator pair, and ¤: shows a stable pair.
and hence, each LNARC design will actually have a wider robust-stability region. It
turns out that for this example, LNARC#1 maintains stability for all k1 ∈ [0.69, 1.75],
LNARC#2 for all k1 ∈ [0.43, 1.58], LNARC#3 for all k1 ∈ [0.25, 1.01], and LNARC#4
for all k1 ∈ [0.25, 0.73]. Of course, performance-robustness is only guaranteed for the
subintervals defined in Table 6.1; see Figure 6.14 for the stability-mismatch cases in
more detail.
For this example, we now follow the process to design the four KFs in the RMMAC
architecture of Figure 6.9.
As explained in Section 5.3, a great deal of care must be exercised in designing the
KFs in the RMMAC architecture, since the convergence of the appropriate posterior
probability to its nearest probabilistic neighbor is at the heart of the RMMAC identi-
fication process (see Chapter 3). The basic decision is how to select the nominal value
∗
of the uncertain spring stiffness, k1 , denoted by k1i , i = 1, 2, 3, 4, for designing each
of the four KFs. We stress that these nominal values are not at the centers of the sets
164
Ωj defined in Table 6.1. Rather, as explained in Section 5.3, the KF design points,
∗
k1i , i = 1, 2, 3, 4, must be determined by an optimization process, so that the Baram
Proximity Measures (BPMs) agree at the boundaries of the sets, or models, Ωj . The
outcome of this optimization process is shown in Figure 6.15 using the optimized KF
nominal values.
Model#1
Model#2
Model#3
Model#4
−8.13
0.25 0.5 0.75 1 1.25 1.5 1.75
k1
Figure 6.15: Baram Proximity Measures (BPMs) for the optimized four KFs (LFD
case). The ♦ show the nominal spring constant used to design the associated KF, see
eq. (6.22).
In Figure 6.15 the curves show the BPM for any value 0.25 ≤ k1 ≤ 1.75 from
each of the four optimized KFs. This process is required so that for any k1 ∈ Ωj , the
corresponding posterior probability Pj → 1, Pk → 0 ∀k 6= j. The specific nominal
KF numerical values for this example are obtained by iteration and are
∗
where k1i = K1KF (i, 1) is the nominal spring constant k1 used in Model#i associated
with the KF#i. In this case, these occur at the minimum of the Li BPM curve over
the corresponding model.
Recall that the GNARC and the LNARC compensators are designed in continuous
165
time. The KFs are designed in discrete-time using a sampling interval Ts . In addi-
tion, the correct variances of the discrete-time white noise sequences, ξ(·) and θ(·),
are calculated and used to design the four KFs and the posterior probability evalua-
tor (PPE); these discrete-time numerical values were used in all Monte Carlo (MC)
simulations in the sequel. Simulations are implemented in Simulink in discrete-time
using a zero-order hold with a sampling time of Ts = 0.01 secs.
As explained in Section 5.4, the real-time KF residual sequences of RMMAC in
Figure 6.9, rj (t); j = 1, . . . , 4, t = 0, 1, 2, . . ., are used by the PPE to generate on-line
the four posterior probabilities, Pj (t); j = 1, . . . , 4, t = 1, 2, . . ., which are next used
to generate the overall RMMAC control signal u(t) by probabilistic weighting, i.e.
4
X
u(t) = Pj (t) uj (t) (6.23)
j=1
where the uj (t) are the “local” controls generated by each LNARC, Kj (s), as designed
in Section 6.3.2.
Now we are ready to construct the RMMAC and simulate it in different cases.
In the sequel, for simulation purposes, the RMMAC is tested and evaluated in three
different cases i.e. under normal conditions, mild, and severe violations of the theoret-
ical assumptions. We used an actual unmodeled time-delay τ of 0.01 seconds inserted
at the control channel2 . (Recall that the maximum time-delay was τ = 0.05 secs.)
In all stochastic simulations, the continuous plant white noise ξ(t) has unit inten-
sity, Ξ = 1. Also, the measurement is assumed to be corrupted by a zero-mean white
Gaussian noise process θ(t), of continuous intensity Θ = 10−6 , and independent of
ξ(t).
The continuous plant is first discretized using zero-order hold with sample time
Ts = 0.01 secs, and the discrete equivalent noise intensities, Q and R, respectively for
the continuous-time noise intensities, Ξ and Θ, are calculated as
Z Ts
T
Q= eAτ LΞLT eA τ dτ
0 (6.24)
R = Θ/Ts
2
Loosely speaking, this time-delay behaves like an unmodeled non-minimum phase zero at s =
200.
166
where Ts is the sampling time used in the discretization (Ts = 0.01 secs).
The integral is computed using the matrix exponential formulas in [126]. However,
if the sampling time Ts is selected small enough, the integral in eq. (6.24) can be
approximately replaced by
Interested readers can refer to, e.g., [94] for more details.
In the sequel some representative3 stochastic simulations are shown using the complete
RMMAC under normal conditions. It is stressed that unless stated otherwise:
(a) all simulations use a stochastic disturbance and white measurement noise gener-
ated according to eq. (6.13). The true system includes an actual (but unmodeled)
time-delay of 0.01 secs in the control channel.
(b) all initial model probabilities are initialized to be Pk (0) = 0.25 ; ∀k = 1, . . . , 4, at
t = 0 sec.
(c) we present numerical averages for 5 MC simulations.
“Easy” identification, I:
The dynamic evolution of the four posterior probabilities when the true k1 = 1.65,
well inside the Model#1, and the corresponding outputs for the RMMAC and the
GNARC systems are shown in Figure 6.16. The correct model (Model#1) is identified
quickly in about 2 secs. The improvement in disturbance-rejection by the RMMAC
vis-a-vis the GNARC is evident as shown in Figure 6.16(b).
3
We have actually done hundreds of such stochastic simulations.
167
P1(t)
0.5
0
1
P2(t) 0.5
0
1
P3(t)
0.5
0
0.4
P4(t)
0.2
0
0 2 4 6 8 10
t(sec)
0.06
GNARC
RMMAC
0.04
0.02
z(t)
−0.02
−0.04
−0.06
0 20 40 60 80 100
time (sec)
Figure 6.16: “Easy” identification I: Simulation results for k1 = 1.65 (LFD case).
“Harder” identification:
When the actual spring constant is near the boundary between two models, it
168
P1(t)
0.5
0
0 5 10 15 20 25 30
1
P2(t)
0.5
0
0 5 10 15 20 25 30
1
P3(t)
0.5
0
0 5 10 15 20 25 30
1
P4(t)
0.5
0
0 5 10 15 20 25 30
t(sec)
0.1
GNARC
RMMAC
0.05
z(t)
−0.05
−0.1
0 20 40 60 80
time (sec)
Figure 6.17: “Easy” identification II: Simulation results for k1 = 0.3 (LFD case).
takes longer (more data) to resolve the true hypothesis. In this example, k1 = 0.405
is selected which belongs to Model #3 but is also “close” to Model #4, see Figure 6.15.
The probabilities vs time as well as output comparisons are shown in Figure 6.18. It
takes about 50 secs to resolve the ambiguity between Models#3 and #4.
Figure 6.18(a) shows that the probabilistic weighing of the control according to
eq. (6.23) persists for about 50 secs. However, as evidenced by Figure 6.18(b), there is
no significant degradation of the RMMAC performance as compared to the GNARC.
169
1
P1(t)
0.5
0
0 20 40 60 80
1
P2(t)
0.5
0
0 20 40 60 80
1
P3(t)
0.5
0
0 20 40 60 80
1
P4(t)
0.5
0
0 20 40 60 80
t (sec)
0.04
GNARC
RMMAC
0.02
z(t)
−0.02
−0.04
0 20 40 60 80 100
t (sec)
Figure 6.18: “Harder” identification: Simulation results for k1 = 0.405 in M#3 (LFD
case).
170
Mismatch instability:
In Table 6.2 the mismatch-stability properties of the LNARC designs were summa-
rized. We evaluate the RMMAC response when we force it to be unstable at t = 0.
Figure 6.19 illustrates a typical result selected from several different MC simula-
tions. In Figure 6.19 the true value of k1 is 1.75 in Model#1; its nearest probabilistic
neighbor is KF#1. From Table 6.2 we know that if we use LNARC#4, K4 (s), with
Model#1 we have an unstable closed-loop system. To force this initial instability at
time t = 0, the initial values of the probabilities are selected to be:
1
P (t)
0.5
1
0
0 1 2 3 4 5
1
P (t)
0.5
2
0.2
0 RMMAC
0 1 2 3 4 5 0.15 GNARC
1
P (t)
0.5
3
0.05
z(t)
0
0 1 2 3 4 5
1
−0.05
P (t)
0.5
4
0
0 1 2 3 4 5
−0.15
t(sec) 0 20 40 60 80 100
time (sec)
Figure 6.19: Mismatch instability. The RMMAC recovers from an initial unstable
configuration at t = 0 (LFD case).
Figure 6.19(a) shows that the “correct probability” P1 (t) → 1 within 1.3 secs,
starting from its initial value P1 (0) = 0.01; the other three probabilities converge
to zero within 1.3 secs as well. Figure 6.19(b) shows the output response in which,
after an initial period of brief “instability”, the RMMAC recovers and returns to its
predictable superior disturbance-rejection.
Figure 6.20 illustrates another mismatch-instability result, similar to the above
171
case with k1 = 1.75. To force this instability the posterior probabilities at time T =
60 secs are forced to be
P1 (T ) = P2 (T ) = P3 (T ) = 0.01, P4 (T ) = 0.97
so that the RMMAC system is forced to be unstable at time t = T for 0.1 secs4 .
Figure 6.20 shows that the RMMAC rapidly recovers to a stable expected configura-
tion. Figure 6.20(a) shows that the “correct probability” P1 (t) → 1 has been forced
1
P1(t)
0.5
0
0 50 100 150
1
P2(t)
0.5
0
0 50 100 150
1
P3(t)
0.5
0
0 50 100 150
1
P4(t)
0.5
0
0 50 100 150
time (sec)
0.06
RMMAC GNARC
0.04
0.02
z(t)
−0.02
−0.04
−0.06
0 50 100 150
time (sec)
Figure 6.20: The RMMAC recovers from a forced unstable configuration at t = 60 secs
lasting for 0.1 secs (LFD case).
4
Such a test was suggested by Prof. Brian D. O. Anderson.
172
to P1 (60) = 0.01. Figure 6.20(b) shows the output response in which the RMMAC
quickly recovers and returns to its predictable superior disturbance-rejection.
Important Remark: Provided that all convergence assumptions hold, [44], we ob-
served in all our simulations that a short temporary instability of the RMMAC will
cause all signals to grow. As a consequence, the signal-to-noise ratio will increase, and
it will be reflected in the size of the residuals; this, in turn, appears to force the PPE
to adjust the posterior probabilities so that a stable configuration is re-established.
In this section, we evaluate the RMMAC performance when some of the theoretical
assumptions are mildly violated. The theory which guarantees the convergence of
the posterior probabilities [44, 45] assumes that all MMAE signals are stochastic
stationary random processes. With some additional ergodicity conditions, all results
presented in Section 6.3.5.1 satisfied those assumptions, and we have indeed observed
convergence to a nearest probabilistic neighbor. In all simulations done under mild
violation of the theoretical assumptions the RMMAC worked well and no instabilities
were observed.
We shall show that the RMMAC still performs well; in a sense it appears “ro-
bust” to mildly violating the theoretical assumptions. We evaluated the RMMAC
performance over a wide variety of operating conditions, for different values of the
uncertain spring stiffness. The following representative cases are presented.
P1(t)
0.5
2
0
1
P2(t)
1 0.5
0
1
d(t)
P3(t)
0.5
0
−1
1
P4(t)
0.5
−2 0
0 60 120 180 240 300 350 0 2 4 6 8 10
time (sec) time (sec)
0.25
RMMAC
0.2 GNARC
0.1
z(t)
−0.1
−0.2
0 60 120 180 240 300 350
time (sec)
(c) Output response
Figure 6.21: RMMAC performance when the actual disturbance d(t) is a ±2 periodic
(filtered) square-wave (LFD case).
Note that the probabilities converge to the correct Model#1 quickly, i.e. P1 (t) →
1 as shown in Figure 6.21(b). After the probability P1 (t) converges to unity, the
RMMAC exhibits a superior performance compared to the GNARC, as shown in
Figure 6.21(c).
1
P (t)
0.5
1
0
0 5 10 15
1
P (t)
0.5
2
0
0 5 10 15
1
P (t)
0.5
3
0
0 5 10 15
1
P (t)
0.5
4
0
0 5 10 15
time (sec)
0
u(t)
z(t)
−0.1 −5
−0.2 −10
0 20 40 60 80 0 5 10 15
time (sec) time (sec)
as shown in Figure 6.23(a). The dashed-lines indicate the times that the spring
crosses the boundaries of the four models of Table 6.1. Figure 6.23(b) shows the dy-
namic evolution of the four posterior probabilities, while Figure 6.23(c) compares the
GNARC and RMMAC performance output responses, which shows that the RMMAC
continues to work quite well.
It is tempting to interpret Figure 6.23(b) as demonstrating a “transient” in the
identification process. This may well be true, but the reader should realize that the
signals generated by the time-varying plant should not be interpreted using “fixed
model” reasoning. After all changing the spring stiffness according to eq. (6.26)
implies an exogenous energy transfer as a function of time, which is not accounted for
in a “fixed parameter” model. This opens up avenues for future research. Nonetheless,
the results of Figure 6.23 (and many others not presented herein) demonstrate the
ability of the RMMAC in dealing with slowly-varying uncertain parameters.
Up to now we found that the RMMAC feedback system works very well when the
theoretical assumptions are not violated. The RMMAC feedback system also works
176
1.75
M#1
1.02
M#2
0.64
M#3
M#4 0.40
0.25
0 200 400 600 800 1000
t (sec)
1
P1(t)
0.5
0
1
P2(t)
0.5
0
1
P3(t)
0.5
0
1
P4(t)
0.5
0
0 200 400 600 800 1000
time (sec)
0.15
GNARC
0.1 RMMAC
0.05
z(t)
−0.05
−0.1
0 200 400 t (sec) 600 800 1000
Figure 6.23: RMMAC responses for the slow sinusoidal parameter variations, as in
eq. (6.26) (LFD case).
177
quite well when some of the theoretical assumptions are mildly violated. However, we
shall see that the RMMAC feedback system may perform poorly, or even break into
instability, when the theoretical assumptions are severely violated. In these cases,
we have isolated the reasons that cause poor performance. Chapter 10 suggests some
ways to modify the RMMAC design to avoid such behavior. including an introduction
of a “fail-safe” safety-net mechanism.
CASE 1A. In this case, Ξact = 100 is used in the simulations. Recall that the
value of design intensity is Ξ = 1. The true value of the spring constant is k1 = 1.5.
The simulation results are shown in Figure 6.24.
The transients of the posterior probabilities are shown in Figure 6.24(a). It can
be seen that there is a “model selection confusion” among the posterior probabilities.
Figure 6.24(b) compares the RMMAC output with the GNARC. Since the posterior
probabilities do not converge they cause the RMMAC to have poorer performance
than the GNARC. However, the closed-loop system remains stable. In fact, as it is
obvious from Figure 6.24(a), the RMMAC most of the time selects the Model#1,
and hence places the compensator LNARC#1 in the feedback loop, but its efforts
178
1 2
P1(t)
0.5 RMMAC
1.5 GNARC
0
1 1
P2(t)
0.5
0 0.5
z(t)
1 0
P3(t)
0.5
0 −0.5
1 −1
P4(t)
0.5
0 −1.5
0 50 100 150 200 0 50 100 150 200
time (sec) time (sec)
(a) Posterior probabilities (b) The RMMAC and the GNARC output z(t)
Figure 6.24: The RMMAC evaluation under severe assumption violation. The plant
disturbance intensity is unit (Ξ = 1) in designing the KFs, but in the simulations it
is chosen Ξact = 100. The value of the spring constant is k1 = 1.5. Only one MC run
is shown. Note that the RMMAC exhibits poor performance in comparison to the
GNARC, due to the frequent switching of the posterior probabilities (LFD case).
to keep the system hedging in this model is not completely successful. It is due to
the fact that such unknown (and large) plant noise intensity will harshly violate the
assumptions underlying of the RMMAC.
It is again stressed that since the design KFs plant noise intensity is not changed,
the BPM is the same as in Figure 6.15 as well as the Kalman Filters are the same as
in eq. (6.22).
1 0.06
P (t)
0.5 RMMAC
1
0 GNARC
0.04
1
P (t)
0.5 0.02
2
z(t)
0
1
P (t)
0.5
3
−0.02
0
1 −0.04
P (t)
0.5
4
0 −0.06
0 25 50 0 20 40 60 80 100
time (sec) time (sec)
Figure 6.25: The RMMAC performance evaluation under severe assumption violation
subject to unknown plant disturbance. The designing plant disturbance intensity is
unit (Ξ = 1) but it is very small in the simulations, Ξact = 0.01. The value of the
spring stiffness is k1 = 0.9 (in Model#2). Note that the “wrong” model, Model#1,
is identified. Only one MC run is shown (LFD case).
due to the fact that the persistent excitation in the model estimation part of the
RMMAC, which is needed for an appropriate identifiability, is not adequate.
Figure 6.25(b) shows the RMMAC output. No instability is observed because the
compensator LNARC#1 can still stabilize the plant with k1 = 0.9; see Table 6.2, and
the RMMAC works pretty well.
8
100
6
4
2
d(t)
Ξ(t)
0
−2
−4
−6
1 −8
0 50 150 250 350 400 0 100 200 300 400
time (sec) time (sec)
(a) Square-wave intensity of the plant distur- (b) Plant disturbance noise
bance
1 2
P1(t)
0.5 RMMAC
0 1.5 GNARC
1
P2(t)
0.5 1
0
z(t)
0.5
1
P3(t)
0.5 0
0
1 −0.5
P4(t)
0.5
0 −1
0 100 200 300 400 0 100 200 300 400
time (sec) time (sec)
Figure 6.26: The RMMAC performance evaluation under temporary severe violation.
The designing plant disturbance intensity is unit (Ξd = 1) but it is a periodic square-
wave in the simulations (Ξact ∈ {1, 100}). The value of the spring stiffness is k1 = 0.9.
Only one MC run is shown (LFD case).
intensity switches to Ξact = 100; Figure 6.26(d) compares the RMMAC output re-
sponses with that of the GNARC. The RMMAC does not yield superior performance
dealing with such uncertain plant noise intensity.
Long-term mismatch-instability:
This case illustrates another type of mismatch-instability result where the RMMAC
is forced to an unstable mismatch configuration at time 48.5 ≤ t ≤ 51.5 secs. The
true value of k1 = 1.5 is used. We want to see if this forced long-term mismatch will
181
The simulation results for a single Monte Carlo run are shown in Figure 6.27.
The transients of the posterior probabilities are shown in Figure 6.27(a) where a
probability confusion starting at t = 48.5 secs is apparent and that the RMMAC
eventually recovers in about 90 secs. Figure 6.27(b) shows the output of the RMMAC
1 0.2
P1(t)
0.5
0
0.1
1
P2(t)
0.5
0
z(t)
0
1
P3(t)
0.5
0 −0.1
1 RMMAC
P4(t)
0.5 GNARC
0 −0.2
0 50 100 150 200 0 50 100 150 200
time (sec) time (sec)
0.05
r(t)
−0.05
−0.1
0 50 100 150 200
time (sec)
Figure 6.27: The RMMAC evaluation under long-term forced instability (48.5 ≤ t ≤
51.5) for k1 = 1.5 in Model#1. Only one MC run is shown. However, the probability
of such an event due to “abnormal” plant disturbance and sensor noise signals is truly
infinitesimal (LFD case). Note that the RMMAC recovers after 170 secs.
182
and that it is temporarily unstable due to this forced instability. The four residual sig-
nals shown in Figure 6.27(c) demonstrate that the residual signals are vastly increased
compared to their computed covariances for about 50 secs (5000 measurements) lead-
ing to the probabilities oscillating. If the system is operating under more or less
normal conditions, the probability that abnormal signals occurring at the “tails” of
the probability density functions, for 5000 time periods, is truly infinitesimal.
We also tested longer mismatched instabilities, e.g. for 6 secs, in which the RM-
MAC could not recover and the result was an unstable configuration. However,
choosing a very large mismatch-instability, such as 6 secs, is completely unrealis-
tic. In summary, severe violations can lead to “confusion” in identification, mediocre
performance, and even instability under intentional forced long-term mismatched-
instabilities.
P1(t)
1.8 0.5
0
1.4 1
P2(t)
0.5
0
k1(t)
1
1
P3(t)
0.5
0.6 0
1
P4(t)
0.2 0.5
0
0 50 100 150 200 0 50 100 150 200
Time (sec) time (sec)
RMMAC
0.8
GNARC
0.6
−1
0 50 100 150 200
time (sec)
The stochastic simulation results presented for the low-frequency design (LFD) demon-
strate the excellent performance of the RMMAC system. Compared to the “best”
non-adaptive design, GNARC, the RMMAC consistently had superior disturbance-
rejection with minimal increase in control activity. Under normal conditions, i.e.
when the theoretical assumptions are not violated, no closed-loop instabilities were
noted in thousands of MC simulations.
The example illustrated how to predict (and then validate) the potential perfor-
mance characteristics of the RMMAC. This was done by analyzing the LTI feedback
loops involving the GNARC and each of the four LNARCs so as to generate RMS
predictions, frequency-domain visualizations etc. We emphasize that this construc-
tive capability is critical so that quantitative tradeoffs can be carried out, before one
constructs and tests the full-blown RMMAC design with numerous MC simulations.
184
In this section, the RMMAC performance was also evaluated when some of the as-
sumptions are “mildly violated”. It was shown that the RMMAC still performs very
well; in a sense it appears “robust” to mild violations of the theoretical assumptions.
The RMMAC was also evaluated when we “severely” violated the assumptions. In
these cases, as to be expected, the RMMAC behaved “poorly” and sometimes prob-
ability oscillations and instabilities were observed. We shall discuss next how to
overcome some of the RMMAC shortcomings, due to unknown plant disturbances,
using the RMMAC/XI architecture discussed in Section 5.7.
Simulation results in Sections 6.3.5.1 and 6.3.5.2 demonstrated that the RMMAC
works quite well even if we violate the theoretical assumptions for MMAE convergence
in a “mild” manner. However, as discussed in Section 6.3.5.3, the RMMAC can behave
poorly when the true stochastic plant disturbances, as quantified by the intensity
matrix Ξact of the actual plant white noise, are very different from those associated
with the intensity Ξd used to design the linear KFs.
If Ξact À Ξd , then the actual residuals are much higher than, say, the 3-sigma val-
ues expected by the residual covariance matrices Sk and this can lead to rapid switch-
ing of the posterior probabilities, hence “confused identification” and poor RMMAC
performance as evidenced in Section 6.3.5.3.
For the sake of simplicity, here we select two intensity values, Ξ1 and Ξ2 , such
that
Ξ1 = 1 , Ξ2 = 100
Then we design two sets of linear KFs, one set indexed by k = 1, 2, 3, 4 using the
noise intensity Ξ1 = 1 and the second set indexed by k = 5, 6, 7, 8 using the noise
intensity Ξ2 = 100. It is important to stress that the nominal points for designing
each KF, as determined by the BPM method of Section 5.3, were different for each
value of Ξ1 and Ξ2 (not shown here).
185
ξ (t); Ξ ∈ {1,100}
θ(t )
u (t ) y (t )
r1 (t ) u1 (t )
K1 ( s )
Designed
r2 (t ) u2 (t )
K 2 ( s)
with
Ξ =1 r3 (t )
u 3 (t )
K 3 ( s)
r4 (t )
u4 (t )
K 4 ( s)
r5 (t )
Designed
r6 (t )
with u (t )
r7 (t ) Dot product
Ξ = 100
r8 (t )
S1 P1 (t )
S2 P2 (t )
S8 P8 (t )
Figure 6.29: The RMMAC/XI architecture for the MSD example (LFD case).
Here, only the simulation results are presented as shown in Figure 6.30. The
square-wave intensity of the plant noise is shown in Figure 6.30(a). The posterior
probabilities are shown in Figure 6.30(b). Recall that only correct probabilities for
k1 = 0.9 ∈ M#2 and M#6, i.e. P2 (t) and P6 (t), are selected that are respectively
associated with the KF#2 (designed with Ξ = 1) and the KF#6 (designed with
Ξ = 100). Other posterior probabilities are almost zero. Figure 6.30(c) compares the
RMMAC and RMMAC/XI outputs where the RMMAC/XI has better performance,
as predicted, than the RMMAC.
186
100
Ξ(t)
1
0 50 150 250 350 400
time (sec)
P (t)
0.5 0.5
1
0 0
1 1
P (t)
P (t)
0.5 0.5
2
0 0
1 1
P (t)
P (t)
0.5 0.5
3
0 0
1 1
P (t)
P (t)
0.5 0.5
4
0 0
0 100 200 300 400 0 100 200 300 400
time (sec) time (sec)
2
RMMAC/XI
1.5 RMMAC
1
z(t)
0.5
−0.5
−1
0 100 200 300 400
time (sec)
Figure 6.30: The RMMAC/XI performance for varying plant noise intensity. The
spring constant k1 = 0.9 is in Model#2 with Ξ = 1 and in Model#6 with Ξ = 100.
Note that the correct probabilities, P2 (t) and P6 (t), (almost immediately) go to unity
as Ξ changes (LFD case).
187
In Section 5.6 we discussed a variant of RMMAC in which only the control associated
with the largest posterior probability at time t is applied to the system.
Figure 6.31 shows a representative comparison between the RMMAC and the
switching version of RMMAC/S with k1 = 1.2 in Model#1. The plant noise and the
sensor noise are white and the normal operating condition is assumed. Only one MC
run is shown. Clearly, the standard RMMAC is better during the initial transient as
compared to the RMMAC/S. This was true for many other simulations (not shown
here). However, as t → ∞, and the posterior probabilities converge, of course both
the RMMAC and RMMAC/S behave in an identical manner.
1 1
P (t)
P (t)
0.5 0.5
1
0 0
1 1
P (t)
P (t)
0.5 0.5
2
0 0
1 1
P (t)
P (t)
0.5 0.5
3
0 0
1 1
P (t)
P (t)
0.5 0.5
4
0 0
0 10 20 30 40 0 10 20 30 40
time (sec) time (sec)
0.05
RMMAC
Switching RMMAC
0.025
z(t)
−0.025
−0.05
0 20 40 60
time (sec)
α = 3 rad/sec (6.28)
2 2
10 10
k =1.75
1 0 k1=1.75
10
Mag.
Mag.
0
10
k =0.25
1 k =0.25
−2 1
10
−2
10
−2 −1 0 1 −2 −1 0 1
10 10 10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)
(a) From disturbance d(t) to output z(t) (b) From control u(t) to output z(t)
Figure 6.32: Bode plot of the open-loop MSD system for different values of spring
constant k1 (HFD case).
189
which reflects our specification for good disturbance-rejection for the frequency
range ω ≤ 3 rad/sec where the disturbance d(t) has most of its power.
The frequency-domain bound on the unmodeled time-delay is the same as in the
LFD case which was shown in Figure 6.3. Also, the following sensor noise weight, and
the frequency-dependent control weight are used in the HFD case in both GNARC
and each LNARC design.
Measurement noise weight: Wn = 10−3
10(s + 300) (6.31)
Control weight: Wu (s) =
s + 3 × 104
The same sensor noise weight (constant) is used as in the LFD case; see eq. (6.14).
Besides, in both designs, the high-pass control weight is selected in such a way that
its zero is two decades above the desired bandwidth (3 rad/sec) and its pole is two
decades further as well.
Figure 6.33(a) shows the low-pass 1st order HFD disturbance modeling filter.
Figure 6.33(b) shows the frequency response of the used control weight.
1
0
10 10
Mag.
Mag.
−1 0
10 10
−2 −1
10 10
−2 −1 0 1 2 0 2 4
10 10 10 10 10 10 10 10
Freq (rad/sec) Freq (rad/sec)
(a) The 1st order disturbance modeling low- (b) Control weight Wu (s), a high-pass fil-
pass filter, Wd (s) ter that penalizes the system for using large
controls at high frequencies
AG
p max = 0.79 (with µub ≈ 0.995) (6.32)
190
12
10
8 FNARC
Ap
4
GNARC
2
0
0.25 0.5 0.75 1 1.25 1.5 1.75
k1
Figure 6.34: Best GNARC and FNARC performance parameters for the HFD calcu-
lated for all k1 ∈ Ω = [0.25, 1.75].
Following the search algorithm, described in Chapter 5, to find the best “models”
and compensators LNARCs using X% = 66% of the FNARC, we obtain Figure 6.35
that shows the results.
12
10
FNARC
8
Ap
4 66% of FNARC
GNARC
2
0
0.25 0.5 0.75 1 1.25 1.5 1.75
k
1
Figure 6.35: Best possible HFD performance parameters, using DGK algorithm, re-
quiring 66% of the best FNARC performance.
5
Compare Figures 6.34 and 6.6 to note the drastic change associated with the HFD case vis-a-vis
the LFD one.
191
The outcome of Figure 6.35 is summarized in Table 6.3. As a result, the large
uncertainty set Ω of eq. (6.5) for the HFD yields seven “models” and seven local com-
pensators LNARCs. Clearly, the reduction in parameter uncertainty allows larger per-
formance gains Ap for designing the LNARCs, resulting into guaranteed both stability-
and -performance robustness over the corresponding subintervals of Table 6.3.
As a result, the GNARC and seven LNARCs, K1 (s), . . . , K7 (s) are designed for
each subinterval shown in Table 6.3. These LNARCs guarantee 66% of the FNARC
performance and a performance improvement of 2.2 to 9 times better than that of
the best GNARC design. It is emphasized that if we require better performance than
66%, we will need more than seven local compensators, LNARCs.
Figure 6.36 compares the frequency response of the GNARC with the seven
LNARCs. Note that at frequencies up to approximately 1 r/s, the LNARCs gen-
erate a loop-gain about 3.3 to 11 times larger than that of the GNARC. Note that
each individual LNARC in the closed-loop design has guaranteed performance- and
stability-robustness over its associated parameter subinterval in Table 6.3.
In order to understand how one can easily predict the potential RMMAC performance
characteristics, assume that one of the posterior probabilities converges to its nearest
probabilistic neighbor; it follows that a specific LNARC is used. After the probability
192
4
10
3
10
2
10 K1(s)
Mag.
1
10
0
10
GNARC
K7(s) LNARC
−1
10
−2 0 2 4
10 10 10 10
Freq. (rad/sec)
z(s)
Disturbance-rejection transfer function: Mdz (s) ≡ (6.33)
d(s)
for different values of the unknown spring constant, k1 ∈ [0.25, 1.75], for the GNARC
and for each LNARC design.
Figure 6.37 illustrates the above using the (true) nominal values of KFs,6 see
eq. (6.34), quantifying the potential RMMAC improvement in disturbance-rejection.
It is stressed that these results are consistent with 1/Ap (as given in Table 6.3) at
low frequencies. Similar plots, such as control signal characteristics, can be made for
other values of the uncertain spring constant.
Figure 6.37 predicts that the RMMAC has the potential to significantly improve
disturbance-rejection; these predictions will be validated in the sequel. Other transfer
functions could also be computed (not shown) from disturbance ξ(t) and the sensor
noise θ(t) to the control u(t).
Figure 6.38 evaluates the potential performance improvement of the RMMAC
using stochastic metrics, namely by comparing the RMS errors of the output z and
6
These are calculated in the next subsection.
193
GNARC GNARC
0 LNARC#1 0 LNARC#2
10 10
Mag.
Mag.
−2 −2
10 10
−2 −1 0 1 2 −2 −1 0 1 2
10 10 10 10 10 10 10 10 10 10
Freq. (rad/sec) Freq. (rad/sec)
GNARC GNARC
0 LNARC#3 0 LNARC#4
10 10
Mag.
−2 Mag. −2
10 10
−2 −1 0 1 2 −2 −1 0 1 2
10 10 10 10 10 10 10 10 10 10
Freq. (rad/sec) Freq. (rad/sec)
GNARC GNARC
0 LNARC#5 0 LNARC#6
10 10
Mag.
Mag.
−2 −2
10 10
−2 −1 0 1 2 −2 −1 0 1 2
10 10 10 10 10 10 10 10 10 10
Freq. (rad/sec) Freq. (rad/sec)
GNARC
0 LNARC#7
10
Mag.
−2
10
−2 −1 0 1 2
10 10 10 10 10
Freq. (rad/sec)
(g) k1 =0.3
0.4
0.2
0
0.2 0.4 0.6 0.8 1
(a) 1.2 1.4 1.6 1.8
RMS from ξ to control u(t)
18
GNARC
16 LNARC
RMS
14
12
10
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
(b)
Uncertain k 1
Figure 6.38: Predicted potential RMS performance of the RMMAC vs GNARC due to
xi(t) only (HFD case). (a) Output RMS comparisons, (b) Control RMS comparisons
the control u, for different values of k1 due to ξ(t) only. They are in fact the RMS
comparisons of GNARC with LNARCs, in the closed-loop transfer function from
disturbance input ξ to output performance and to control signal. These RMS results
are readily computed by solving standard covariance algebraic Lyapunov equations
for stochastic LTI systems [56]. The graphs of Figure 6.38(a) clearly suggest that
RMMAC has the potential of decreasing the output RMS by a factor of 1.7 to 4.8
over the GNARC system.
The different HFD “stability-mismatch” cases are shown in Table 6.4.
Table 6.4, similar to the LFD case, predicts the closed-loop stability if the LNARC#j,
Kj (s) ; (j = 1, . . . , 7) is used when the true model is the Model#i (i = 1, . . . , 7), i.e.
true spring constant is in subinterval#i.
is used here; see Section 3.5. Thus, the seven models used for KFs are designed based
on the following nominal points.
where K1KF (i, 1) is the nominal coefficient spring k1 used in Model#i associated
with the KF#i.
Figure 6.39 shows the fact that the BPMs agree at the model boundaries, using
the values of eq. (6.34).
−5.82
−5.84
BPM
−5.86 L7 L6 L L4
5 L3 L2 L1
−5.88
0.4 0.6 0.8 1 1.2 1.4 1.6
k1
Figure 6.39: BPM for the HFD. The nominal k1 of the KFs (not shown) are at the
minimum values of the Li (BPM) curves; see eq. (6.34).
196
After designing the LNARCs and KFs, the RMMAC can be implemented and eval-
uated for different conditions, namely under normal conditions, mild violations of
assumptions, and severe violations of assumptions. The simulation was implemented
in discrete time using a zero-order hold with a sampling time of Ts = 0.001 secs7 .
In addition, the correct variances of the discrete-time white noise sequences, ξ(·) and
θ(·), were calculated and used to design the seven KFs and the posterior probability
evaluator (PPE); these discrete-time numerical values were used in all Monte Carlo
(MC) simulations in the sequel.
Among the numerous simulations done, only some representative results of RM-
MAC stochastic responses under normal conditions are shown here. Unless stated
otherwise the results are averaged over 5 Monte Carlo (MC) runs.
The dynamic behavior of the seven posterior probabilities when the true k1 =
1.65, well inside the subinterval (model) #1, and the corresponding outputs for the
RMMAC and the GNARC systems are shown in Figure 6.40. The correct model
(Model#1) is identified in about 40 secs. The improvement in disturbance-rejection
by the RMMAC is evident as shown in Figure 6.40(b).
Figure 6.41 shows a different simulation when the true k1 = 0.45, inside the
subinterval#6 but close to subinterval#7. The correct model (Model#6) is identified
in about 100 secs. Note that the RMMAC takes longer to resolve the ambiguities
between models #6 and #7, because k1 is very close to the boundaries of models #6
and #7.
Monte Carlo results of probabilities ; k1=1.65 Monte Carlo results of probabilities ; k1=0.45
1 1
P1(t)
P1(t)
0.5 0.5
0 0
1 1
P2(t)
P2(t)
0.5 0.5
0 0
1 1
P3(t)
P3(t)
0.5 0.5
0 0
1 1
P4(t)
P4(t)
0.5 0.5
0 0
1 1
P5(t)
P5(t)
0.5 0.5
0 0
1 1
P6(t)
P6(t)
0.5 0.5
0 0
1 1
P7(t)
P7(t)
0.5 0.5
0 0
0 10 20 30 40 50 60 70 80 90 100 0 50 100 150 200 250 300 350
t(sec) t(sec)
Monte Carlo results of performance output ; k1=1.65 Monte Carlo results of performance output ; k1=0.45
0.6 0.6
RMMAC RMMAC
GNARC GNARC
0.4 0.4
0.2 0.2
0 0
Z(t)
Z(t)
−0.2 −0.2
−0.4 −0.4
−0.6 −0.6
−0.8 −0.8
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
time (sec) time (sec)
Monte Carlo results of control signal ; k1=1.65 Monte Carlo results of control signal ; k1=0.45
100 100
RMMAC RMMAC
GNARC GNARC
50 50
0 0
Z(t)
Z(t)
−50 −50
−100 −100
−150 −150
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.25 0.5 0.75 1
time (sec) time (sec)
Figure 6.40: Monte Carlo stochastic simu- Figure 6.41: Monte Carlo stochastic simu-
lation results when true spring constant is lation results when true spring constant is
k1 = 1.65 (HFD case). k1 = 0.45 (HFD case).
198
In this example, the true spring value is k1 = 0.635 in Model#3 and also close
to Model#2 as one could expect a long period of time for the posterior probabilities
to converge. Thus, we shall see the difference between the probabilistically blended
control signals (due to the LNARC#3 and LNARC#2) in the RMMAC and the
perfect control signal of LNARC#3.
In Figure 6.42 we compare the performance of the RMMAC obtained with k1 =
0.635 against the performance obtained when we use the perfect compensator LNARC#3
only, i.e. we fix P3 (t) = 1 and set P1 (t) = P2 (t) = P4 (t) = 0 ∀t = 0, 1, 2, . . .. Ob-
viously, in this case we have “perfect” identification from the start. Figure 6.42(a)
shows the posterior probabilities. Recall that the probabilistic averaging results in
insignificant performance deterioration even if the convergence of the posterior proba-
bilities happened only after about 120 secs. Figure 6.42(b) shows that the probabilistic
averaging of the controls are not so different than the perfect local control namely
u3 (t). The performance response of the perfect compensator is slightly better than
that of the RMMAC in the transient period as shown in Figure 6.42(c). However, af-
ter t = 120 secs and the posterior probabilities converge, naturally both the RMMAC
and the perfect LNARC#3 behave in an identical manner.
Mismatch instability:
We next evaluate the RMMAC response when it is forced to be unstable at time
t = 0. Recall that Table 6.4 summarized the mismatch-stability cases of the LNARC
designs.
Figure 6.43 illustrates a typical result selected among several of MC simulations.
In Figure 6.43 the true value of k1 is 0.3; its nearest probabilistic neighbor is Model#7.
Referring to Table 6.4, the compensator LNARC#1, K1 (s), with Model#7 corre-
sponds to an unstable closed-loop system. To impose this forced instability, the
initial values of the probability vector are selected to be
1 3
P1(t)
0.5 2
0
1 1
P2(t)
0.5
0
u(t)
0
1 −1
P3(t)
0.5
−2
0
1 −3 LNARC#3
P4(t)
0.5 RMMAC
0 −4
0 50 100 150 0 50 100 150
time (sec) time (sec)
× 10−2
1.5
LNARC#3
RMMAC
0.75
z(t)
−0.75
−1.5
0 50 100 150
time (sec)
Figure 6.42: Performance comparison of the LNARC#1 with the RMMAC for k1 =
0.635 (HFD case). Numerical averages for 5 MC simulations are presented and there
is no sensor noise.
as depicted in Figure 6.43, the RMMAC recovers to a stable (and correct) configu-
ration. Figure 6.43(a) shows that the “correct probability” P7 (t) → 1 within about
20 secs, starting from its initial value P7 (0) = 0.01; the other six probabilities converge
to zero within 10 secs as well. Figure 6.43(b) shows the output performance response
in which the RMMAC recovers and returns to its predicted disturbance-rejection.
to the RMMAC convergence in such cases where one model can be stabilized with
more than one local compensator. The main question in such cases is as follows: how
is the RMMAC able to identify the correct model when more than one compensator
can stabilize the unknown plant and to yield the guaranteed robust performance? To
impose this forced mismatch instability the probabilities vector at time t = 0 secs is
P1(t) 1 1
P1(t)
0.5 0.5
0 0
1 1
P2(t)
P2(t)
0.5 0.5
0 0
1 1
P3(t)
P3(t)
0.5 0.5
0 0
1 1
P4(t)
P4(t)
0.5 0.5
0 0
1 1
P5(t)
P5(t)
0.5 0.5
0 0
1 1
P6(t)
P6(t)
0.5 0.5
0 0
1 1
P7(t)
P7(t)
0.5 0.5
0 0
0 20 40 60 0 20 40 60
time (sec) time (sec)
(a) Posterior probabilities of PPE (a) Posterior probabilities of PPE
1.5 1.5
RMMAC RMMAC
GNARC GNARC
1 1
0.5 0.5
z(t)
z(t)
0 0
−0.5 −0.5
−1 −1
−1.5 −1.5
0 25 50 75 100 0 25 50 75 100
time (sec) time (sec)
4
RMMAC
GNARC
2
1
P1(t)
z(t)
0.5 0
0
1
P2(t)
−2
0.5
0
1 −4
P3(t)
0 50 100 150
0.5 time (sec)
0
1
(b) Output performance z(t)
P4(t)
0.5
60
0 RMMAC
1 GNARC
40
P5(t)
0.5
0 20
1
u(t)
0
P6(t)
0.5
0 −20
1
P7(t)
−40
0.5
0 −60
0 50 100 150 0 50 100 150
time (sec) time (sec)
Figure 6.45: Simulation results of the RMMAC performance when the actual distur-
bance d(t) is a ±2 periodic square-wave with the period of T = 60 secs, and k1 = 1.75
(HFD case). Only one MC run is shown.
The KFs in the RMMAC were designed under the assumption that the sensor
noise was white, as in Section 6.4.4.1.
Once more, we performed several MC simulations using different values for the
unknown spring constant. In all cases the RMMAC design worked very well and
203
1.5
RMMAC
GNARC
1
0.5
1
P1(t)
0.5
z(t)
0
0
1 −0.5
P2(t)
0.5
−1
0
1 −1.5
P3(t)
0 25 50 75 100
0.5 time (sec)
0
1 (b) Output performance z(t)
P4(t)
0.5
0 40
RMMAC
1 GNARC
P5(t)
0.5 20
0
1
P6(t)
u(t)
0
0.5
0
1 −20
P7(t)
0.5
0
0 25 50 75 100 −40
0 2 4 6 8 10
time (sec) time (sec)
Figure 6.46: Monte Carlo simulation results of the RMMAC performance for sinu-
soidal sensor noise when k1 = 1.75 (HFD case).
we did not observe any instabilities. Figure 6.46 shows a representative average of
10 MC runs using the value k1 = 1.75 in model#1 – see Table 6.3. An interesting
and instructive case is illustrated in Figs. 6.46(a) and (b), where it takes 20 secs
for the “correct” probability P1 (t) → 1 and for P2 (t) to decay to zero. During this
20 secs time-interval there is significant blending of the “local” control signals u1 (t)
and u2 (t) by the RMMAC (see Figure 6.9). Nevertheless, the RMMAC still delivers
excellent disturbance-rejection (Figure 6.46(b)) with small controls (Figure 6.46(c))
which include the presence of the sinusoidal sensor noise. Thus, control probabilistic
averaging did not produce any serious problems; on the contrary, the RMMAC still
outperforms the GNARC system.
We conclude from the results of this section that, at least for this HFD test case,
204
the RMMAC system still works extremely well, even if we violate some of the key
assumptions; it continues to outperform the GNARC design.
1.75
1.40
1.11 1
k1(t)
P1(t)
0.86 0.5
0
0.66
1
P2(t)
0.49
0.5
0.36
0.25 0
0 200 400 600
1
Time (sec)
P3(t)
0.5
(a) Slow time-varying spring stiffness (sinu- 0
soidal) 1
P4(t)
0.5
2 0
RMMAC
GNARC 1
P5(t)
1 0.5
0
1
z(t)
0
P6(t)
0.5
0
−1 1
P7(t)
0.5
−2 0
0 175 350 525 700 0 175 350 525 700
time (sec) time (sec)
Figure 6.47: The RMMAC performance evaluation with slow sinusoidal k1 (t) variation
(HFD case). Only one MC run is shown.
205
Figure 6.47(a) shows the slowly time-varying spring used in the simulation. Fig-
ure 6.47(b) compares the output of the GNARC with the RMMAC. Recall that the
RMMAC performance is much better at times that the spring constant has larger val-
ues, i.e. in Model#1, as expected in the time-domain RMS predictions; see Figure 6.38
for a similar case in the absence of the sensor noise. Figure 6.47(c) shows the behavior
of the posterior probabilities following the sinusoidal spring of Figure 6.47(a). In many
other simulations (not shown) the RMMAC showed significant disturbance-rejection
compared with the best GNARC under such mild violations, i.e. by introducing
slowly varying springs. It is stressed that these results are very similar to the normal
operating conditions where only fixed parameter are used.
In this section, the RMMAC feedback system is evaluated for different cases when
the theoretical assumptions are severely violated. We shall see that under these
conditions, the RMMAC may perform poorly, or even become unstable.
CASE 1A. In this case, Ξact = 100 is used in the simulations; recall that the
value of the design intensity was Ξ = 1. The true spring constant is k1 = 1.65 inside
the subinterval#1. The simulation results are shown in Figure 6.48.
206
1
P1(t)
0.5
0
1
P2(t)
0.5
0
1
P3(t)
0.5
0
1
P4(t)
0.5
15
0 RMMAC
GNARC
1 10
P5(t)
0.5
5
0
1
z(t)
0
P6(t)
0.5
0 −5
1
−10
P7(t)
0.5
0 −15
0 40 80 120 160 200 0 50 100 150 200
t(sec) time (sec)
Figure 6.48: Case 1A. The RMMAC evaluation under severe assumption violations.
The plant disturbance intensity is unit (Ξ = 1) when designing the KFs, but in the
simulations, we use Ξact = 100 (HFD case). The value of the spring constant is
k1 = 1.65. Only one MC run is shown.
The transients of the posterior probabilities are shown in Figure 6.48(a). It can
be seen that there is a model selection “confusion” among the posterior probabilities.
Figure 6.48(b) compares the RMMAC output with the GNARC. Recall that the poste-
rior probabilities do not converge and oscillate leading the RMMAC to have mediocre
performance. However, the closed-loop system has not gone unstable. Indeed, the
closed-loop system remains stable due to the fact that all LNARCs can stabilize the
plant with k1 = 1.65; see the mismatch cases that were shown in Table 6.4.
CASE 1B. In this case, the plant disturbance intensity is switching between
two different intensities, i.e. Ξact = 1 and Ξact = 100. The KFs design use Ξ = 1 and
the true spring is k1 = 0.3 in Model#7. If, again, the probabilities will oscillate, it is
of interest to see the RMMAC behavior in this case because the mismatch instability
207
table – see Table 6.4 – shows that the compensators LNARC#1, #2, and #3 would
create unstable closed-loop systems with this spring value. The simulation results are
shown in Figure 6.49.
The transients of the posterior probabilities are shown in Figure 6.49(a). It can be
seen that, similar to Case 1A, there is a model selection confusion (oscillation) among
the posterior probabilities due to strong disturbance input. Figure 6.49(b) compares
the RMMAC output with the GNARC. Note that the posterior probabilities do not
converge, when Ξact = 100, similar to Case 1A. In fact such an unknown (and large)
plant noise intensity has harshly violated the assumptions of the RMMAC. Recall
also that the RMMAC shows poor performance in comparison to the GNARC. In the
sequel, we will implement the RMMAC/XI which will fix this RMMAC shortcoming
subject to the uncertain plant noise intensities.
1
100
P1(t)
0.5
0
Ξact(t)
1
P2(t)
0.5
1 0
0 50 100 150 200 250 300 1
time (sec)
P3(t)
0.5
(a) Square-wave intensity of plant noise 0
1
P4(t)
60 0.5
RMMAC
GNARC 0
40 1
P5(t)
0.5
20 0
z(t)
1
P6(t)
0 0.5
0
−20 1
P7(t)
0.5
−40 0
0 50 100 150 200 250 300
0 50 100 150 200 250 300
time (sec)
time (sec)
(c) Performance output z(t)
(b) Posterior probabilities
Figure 6.49: Case 1B. The RMMAC evaluation under severe assumption violations.
The plant disturbance intensity is Ξ = 1 when designing the KFs, but in the simula-
tions it switches in Ξact ∈ {1, 100} (HFD case). The value of the spring constant is
k1 = 0.3. Only one MC run is shown.
208
CASE 1C. In this case, the plant disturbance intensity is, once more, Ξ = 1
in the KFs design, but in the simulations a smaller one, Ξ = 0.01, is used. The true
spring is k1 = 0.4 in Model#6.
The simulation results are shown in Figure 6.50. The transients of the posterior
probabilities are shown in Figure 6.50(a). It can be seen that the probability conver-
gence, in comparison to the normal stochastic case with Ξact = 1 (see Figure 6.41)
is much slower, probably because the persistent excitation due to the smaller plant
disturbance is not adequate.
Figure 6.50(b) shows the RMMAC output. It shows first that the RMMAC is still
better than the GNARC even if the probabilities take long time to settle down, about
130 secs. Second, blending two control signals (LNARC#6 and LNARC#7) can still
0.2
RMMAC
1 GNARC
P1(t)
0.5 0.1
0
1
z(t)
0
P2(t)
0.5
0 −0.1
1
P3(t)
0.5
−0.2
0 0 50 100 150 200
time (sec)
1
P4(t)
0.5 GNARC
0 5
1
P6(t)
u(t)
0.5 0
0
1 −5
P7(t)
0.5
0 −10
0 40 80 120 160 200 0 50 100 150 200
t(sec) time (sec)
Figure 6.50: Case 1C. The RMMAC performance evaluation under severe assumption
violation where the designing plant disturbance intensity is Ξ = 1 but it is very small
in the simulations, Ξact = 0.01 (HFD case). The value of the spring stiffness is
k1 = 0.4 in Model#6. Only one MC run is shown.
209
yield a stable closed-loop during the long transient time in which the probabilities
have not converged. The control signals are compared in Figure 6.50(c).
Figure 6.52(a) shows the behavior of the seven posterior probabilities. Figure 6.52(b)
compares the GNARC and RMMAC performance output responses, which shows that
210
1
P1(t)
0.5
0
1
P2(t)
0.5
0
1
P3(t)
0.5
0
1
P4(t)
0.5
0 RMMAC
2 GNARC
1
P5(t)
0.5
1
0 z(t)
1
P6(t)
0.5 0
0
1 −1
P7(t)
0.5
0 −2
0 50 100 150 200 250 0 50 100 150 200 250
time (sec) time (sec)
Figure 6.51: Forced mismatch instability for 50 ≤ t ≤ 65 secs (HFD case). The true
value is k1 = 0.6 in Model#5. Note the rapid recovery of the RMMAC after the
forced instability, i.e. P5 (t) → 1. Only one MC is shown.
the closed-loop system is not unstable and the RMMAC performance is somewhat
better than the GNARC. Figure 6.52(c) show the control signal of the RMMAC. The
reason that the control signals are shown entirely for 200 secs is indeed to observe the
behavior of the RMMAC control in such a situation in which the posterior probabil-
ities have not converged and the global RMMAC control is a blended control signal
of essentially two LNARC control signals, u6 (t) and u7 (t).
We examined different sinusoidal frequencies, as well as different random seeds, in
an effort to find a case in which the RMMAC shows an oscillatory behavior similar
to the LFD case (see Figure 6.28), or goes unstable, but we were not successful in
this HFD case.
211
P (t) 1
0.5
1
2
0 RMMAC
1 GNARC
P (t)
0.5 1
2
0
1
z(t)
0
P (t)
0.5
3
0
1 −1
P (t)
0.5
4
0 −2
0 50 100 150 200
1
time (sec)
P (t)
0.5
5
0
(b) Output response
1
P (t)
400
0.5
6
200
0
0
u(t)
1
P (t)
0.5 −200
7
0 −400
0 50 100 150 200 0 50 100 150 200
time (sec) time (sec)
The simulation results in Case 1B, similar to the LFD in Section 6.3.5.3, demon-
strate that in the HFD the RMMAC also behaves poorly when the true stochastic
plant disturbances, as quantified by the intensity matrix Ξact of the actual plant
white noise, are very large compared to those associated with the intensity Ξ used
to design the LTI KFs. To overcome this RMMAC shortcoming, we shall use the
RMMAC/XI architecture suggested in Section 5.7. Figure 6.53 shows the simulation
results. The true spring is k1 = 0.3 as in Case 1B. Figure 6.53(a) shows the posterior
probabilities associated with the set#1 of the KFs designed for Ξ = 1. Similarly,
Figure 6.53(b) shows the posterior probabilities associated with the set#2 of the KFs
designed for Ξ = 100. The filtered plant disturbance is driven by two different white
noise intensities switched between Ξact = 1 and Ξact = 100 every 50 secs as shown in
Figure 6.53(c). Figure 6.53(d) compares the output response of both the RMMAC
and GNARC. From the RMS viewpoint, we shall recall that the steady-state RMS
value of the GNARC response in Figure 6.53(d) is 4.2 while for the RMMAC it is 2.6,
i.e. it outperforms the GNARC. We conclude from the results of this section that the
RMMAC/XI architecture “cures” the shortcomings when the plant is subject to un-
known plant disturbances in which we shall consider the plant disturbance intensity
as another uncertain element in the RMMAC/XI method.
Figure 6.54 compares the RMMAC and RMMAC/XI outputs (from Figure 6.49(c)
and 6.53(d)).
213
100
Ξ (t)
act
1
0 50 100 150 200 250 300
time (sec)
1 1
P1(t)
P8(t)
0.5 0.5
0 0
1 1
P2(t)
P (t)
0.5 0.5
9
0 0
1 1
0.5 0.5
10
0 0
1 1
P4(t)
0.5 0.5
11
0 0
1 1
P5(t)
0.5 0.5
12
0 0
1 1
P6(t)
0.5 0.5
13
0 0
1 1
P7(t)
0.5 0.5
14
0 0
0 100 200 300 0 100 200 300
time (sec) time (sec)
40 10
20
0
d(t)
z(t)
0
−10
−20
RMMAC/XI
−40 −20
GNARC
(c) Plant disturbance d(t) (d) The GNARC and the RMMAC/XI outputs
Figure 6.53: The RMMAC/XI performance for the HFD when Ξ is a periodic “square-
wave” as in 6.53(a). The true spring constant is k1 = 0.3 in Model#7 when Ξ = 1
and in Model#14 when Ξ = 100. Note that the “correct” posterior probabilities P7 (t)
and P14 (t) quickly respond to the changes in Ξ in 6.53(a).
214
60
RMMAC
RMMAC/XI
40
20
z(t)
−20
−40
0 50 100 150 200 250 300
time (sec)
Figure 6.54: The RMMAC/XI performance compared to the standard RMMAC sub-
ject to the unknown plant noise intensities. The true spring is k1 = 0.3 (HFD case).
1 1
P1(t)
P (t)
0.5 0.5
1
0 0
1 1
P2(t)
P (t)
0.5 0.5
2
0 0
1 1
P3(t)
P (t)
0.5 0.5
3
0 0
1 1
P4(t)
P (t)
0.5 0.5
4
0 0
1 1
P5(t)
P (t)
0.5 0.5
5
0 0
1 1
P6(t)
P (t)
0.5 0.5
6
0 0
1 1
P7(t)
P (t)
0.5 0.5
7
0 0
0 25 50 75 100 0 25 50 75 100
time (sec) time (sec)
0.5
z(t)
−0.5 RMMAC
RMMAC/S
−1
0 25 50 75 100
time (sec)
(c) The performance output z(t)
Figure 6.55: Performance evaluation of the RMMAC and Switching RMMAC/S. Only
one MC run is shown. The true value of spring constant is k1 = 0.5 in Model#5, but
also close to Model#6 (HFD case).
216
In all simulations up to present, only a single noisy measurement (on x2 (t)) was used.
It is important to find out what might happen to the best (maximum) performance
parameters of both GNARC and the FNARC if we increase the number of measure-
ments. The reason is indeed to find out if we still need any adaptive control and its
potential benefits if more measurements can be made.
Figure 6.56 compares, for the MSD system, the cases where there is only one
measurement on position of mass m2 , x2 (t) with that of the two measurement case in
which another position sensor (an additional measurement) is placed on x1 (t) so that
the position of mass m1 could also be measured, with the same sensor noise eq. (6.3).
It is important to stress that we only show the results of the performance gain
comparisons; we do not intend to use them to design the RMMAC for this MSD ex-
ample. A more complex three-cart MSD system, with two unknown real parameters,
two disturbances, two noisy sensor measurements, two controls and two performance
variables will be analyzed and evaluated in Chapter 9.
Figure 6.56(a) shows a comparison of the performance gains between the one and
two measurement case for the LFD case. Recall that the performance improvement by
using one more measurement associated with the FNARC is between 0–28%, while
for the GNARC it is 84% (see the gap in the Ap -axis between one measurement
and two measurements in logarithmic scale). This result together with many other
experiments show that the GNARC usually has a greater benefit of using additional
measurements than the FNARC for the low-frequency designs.
Figure 6.56(b) shows another comparison of the performance improvement be-
tween one and two measurement case for the HFD case. Here, the performance im-
provement by using one more measurement associated with the FNARC is between
5–15%, while for the GNARC it is 42% which are not as large as the LFD as one
could intuitively expect.
As a result, the author strongly suggests that one must have a quantifiable choice
217
on the degree of performance improvement that one desires from the adaptive system.
In the author’s opinion, the designer must decide on whether or not to implement
an RMMAC with less measurements, or accept the inferior performance and use the
GNARC with more measurements. This allows the designer to make an engineering
trade-off of the complexity vs performance in each case.
2 Meas.
3
10
1 Meas.
FNARC
Ap
10
2 2 Meas.
1 Meas.
GNARC
1
10
0.25 0.75 1.25 1.75
k1
14
12 FNARC
10
8
Ap
s.
ea
6 2M
s.
ea
4 1M GNARC
2
2 Meas.
1 Meas.
0
0.25 0.75 1.25 1.75
k1
Figure 6.56: Comparison of performance gains for the MSD system with one and two
measurements.
218
7.1 Introduction
In this chapter, numerical results for the RMMAC for a SISO non-minimum phase
(NMP) plant are presented. This approach follows the material of Chapter 5 in a
straight-forward manner. The reader should recall [108,125,127,128] that NMP plants
present inherent limitations to performance, e.g. superior disturbance-rejection, espe-
cially if the right-half plane (RHP) zero is close to the jω-axis and good performance
is required in the vicinity of the RHP zero frequencies. Indeed this was one of the
main reasons that we decided to investigate the RMMAC for a NMP plant, as well.
The structure of this chapter is as follows. In Section 7.2 the example of the
non-minimum phase (NMP) zero is described. In Section 7.3 the global and local
compensators for the NMP plant are designed. The bank of Kalman filters needed
in the RMMAC structure are designed in Section 7.4. In Section 7.5 the potential
performance of the RMMAC is compared to that of the GNARC in the course of
frequency domain and the RMS comparisons. The simulation results under various
conditions are shown in Section 7.6. Section 7.7 investigates the trade-off between the
number of models adopted and the performance achieved by the RMMAC. Finally,
conclusions are summarized in Section 7.8.
221
222
x (t)
Wd
q (t)
d (t)
u (t ) + + y (t )
NMP Plant
The Bode magnitude plot of the open-loop transfer function of eq. (7.1) for dif-
ferent values of non-minimum phase zero Z is shown in Figure (7.2). It stands to
reason that this Bode plot (from u(t) to y(t)) is the same as from d(t) to y(t); see
large disturbance amplifications for smaller values of Z.
The plant is subject to a low-frequency stochastic disturbance input d(t) obtained
by filtering continuous-time white noise ξ(t) with zero mean and unit intensity, as
follows.
d(s) α
Wd (s) = = (7.4)
ξ(s) s+α
223
1
10
Z=1
0
10
Magnitude
−1
10
Z = 100
−2
10
−3
10
−2 −1 0 1 2
10 10 10 10 10
Frequency (rad/sec)
Figure 7.2: Frequency response of the open-loop transfer function as in eq. (7.1).
0
10
Magnitude
−1
10
−2
10
−3 −2 −1 0 1
10 10 10 10 10
Frequency (rad/cec)
Figure 7.3: The Bode magnitude plot of the disturbance dynamics as in eq. (7.4).
The state differential equations using the standard controllable form and including
the disturbance dynamics are:
with
0 1 0 0 0 0
0 0 1 0 0 0
A= ; B = ; L =
−45 −49.5 −14.5 1 0
1 (7.6)
0 0 0 −0.1 0 0.1
h i
C = 54 27 − 54Z −1 − 27Z −1 0
Eq. (7.6) demonstrates that the NMP uncertainty is directly translated into the
output matrix C in this state-space representation of the NMP plant. The output
matrix C can be written as
1
C = C0 + C̃ (7.7)
Z
where
h i
C0 = 54 27 0 0
h i
C̃ = 0 − 54 − 27 0
Remark: Note that the RMMAC design can take care of uncertainties in different
parts of the state-space representation. For example, in Chapter 6 a mass-spring-
damper example was investigated in which the spring stiffness uncertainty showed up
in the state matrix, A.
Figure 7.4 shows the NMP plant with weights as required by mixed-µ synthesis.
One can note that there are two frequency-weighted “errors” z̃(t) and ũ(t). This figure
is in fact a block diagram of the uncertain closed-loop NMP system illustrating the
disturbance-rejection performance objective namely the closed-loop transfer function
from ξ(t) → z(t), or d(t) → z(t).
The output error is represented by z̃ that is the main performance variable for
evaluating the quality of the disturbance-rejection. The output errors are most im-
portant below the corner frequency 0.1 rad/sec in our µ-design. The control error is
represented by ũ(t) where the control signal is driven by Wu (s), a high-pass filter to
penalize the system for using large controls at high frequencies.
For evaluating the quality of the disturbance-rejection, the plant output error is
225
ξ (t)
δZ
d (t )
∆un −Z / Z 1/ Z
−Z / Z 1/ Z
x(t ) z (t ) z (t )
∫
θ (t)
u (t )
u (t ) y (t )
Figure 7.4: NMP system with weights to be used in the mixed-µ synthesis.
ω > 103 rad/sec. The control performance objective is to reject (stochastic and/or de-
terministic) disturbances whose power is consistent within the disturbance dynamics
of Figure 7.3.
For the purposes of this example, it is also assumed that there is an unmodeled
time-delay (τ ), representing a dynamic error in the system, but not incorporated in
the state equations given in eq. (7.5). The frequency-domain upper-bound for the
unmodeled time-delay is required for mixed-µ synthesis design and is the magnitude
of the first order transfer function
2.45s
Wun (s) = (7.10)
s + 40
The maximum possible value of τ chosen is τ = 0.05 secs at the plant control
channel input, i.e. τ ≤ 0.05. The frequency behavior of the unmodeled time-delay
including its upper-bound is shown in Figure 7.5(b).
1 1
10 10
Unmodeled time−delay
Frequency bound
0
10
Magnitude
Magnitude
0 −1
10 10
−2
10
−1 −3
10 10
−1 0 1 2 3
−1 0 1 2 3 4
10 10 10 10 10 10 10 10 10 10 10
Frequency (rad/cec) Frequency (rad/cec)
(a) Control signal weight, Wu (s) (b) Unmodeled time-delay dynamic and its
frequency bound, Wun (s)
Figure 7.5: Frequency weights used in the GNARC and each LNARC design.
2
Ap
10
GNARC
FNARC
1
10
0 1 2
10 10 10
Z
Figure 7.6: Best GNARC and FNARC performance gains for the NMP plant. Note
that the scale of both the Ap -axis and the Z-axis is logarithmic.
The GNARC, based on mixed-µ synthesis, is designed using the initial large NMP
zero uncertainty of eq. (7.2).
228
The performance parameter, Ap , is that associated with the plant output y(t).
The GNARC compensator, denoted by K(s), guarantees robust-stability and the best
possible performance for the initial uncertainty interval of eq. (7.2). In order to have
the best possible guaranteed performance using the GNARC, the Ap is maximized
until the mixed-µ upper-bound exceeds unity at some frequency. To this effect, the
performance parameter Ap is maximized until the µ upper-bound is about 0.995.
The maximum performance parameter Ap associated with the GNARC in eq. (7.8)
covering the initial uncertainty bound of eq. (7.2) was determined to be (see Fig-
ure 7.6)
AG
p max = 12.5
By using the performance parameter of eq. (7.3), the best GNARC is then designed
via Matlab. The frequency response and other simulation results using the GNARC
are illustrated later in this chapter.
Following the proposed RMMAC procedure presented in Section 5.4.2, the step-
by-step algorithm there is used to first find the minimum number of the best cov-
ering parameter subsets (models) and then design the “Local Non-Adaptive Robust
Compensators” (LNARCs) based on mixed-µ, each for their corresponding small pa-
rameter subinterval. For each “local” design the same bounds for the unmodeled
time-delay, and the same disturbance, sensor and control weights are used as in the
GNARC. Again, the performance parameter Ap is maximized as much as possible for
each model output, until the corresponding mixed-µ upper-bound is between 0.995
and 1.
The desired level of performance is selected to be X% = 30% of the FNARC
designs as the performance goal of the RMMAC. The minimum number, N, of required
models and their uncertainty range are obtained as shown in Figure 7.7. Figure 7.7
shows that, with the 30% of the FNARC designs requirement, the large parameter
uncertainty interval of eq. (7.2) is subdivided into four subintervals. It also shows
the resulting sequence of the maximized Ap ’s, associated with the plant output y(t),
using the notation introduced in Chapter 5.
The derived covering subintervals for LNARCs are summarized in Table 7.1. In
229
1000
Γ1
Γ2
298
Γ3
Γ4
107
Ap
42.6 Best GNARC
FNARC
26.5 30% of FNARC
Performace levels
12.5
10
1 1.34 2.38 4.1 100
Z
Figure 7.7: “Best” performance gains for each of the four LNARCs using the pro-
posed RMMAC algorithm. The LNARCs are designed for the uncertainty subsets by
considering X% = 30% of the best FNARC performance.
this table, the third column is the best LNARC percentage performance gain (PG%)
vs “FNARC ” used in subsequent designs.
As a result, the interval of uncertainty for the NMP zero, Z, is divided into four
“local” subintervals, obviously non-uniform segments. Also, there are more parameter
intervals near the worst case, Z = 1, required for a robust design. Due to the
inherent NMP zero limitation close to the jω-axis it is reasonable to expect more
models in the adaptive scheme near Z = 1. Table 7.1 indicates that the reduction
in parameter uncertainty allows larger performance parameters Ap for designing the
LNARCs, resulting into guaranteed both stability- and performance-robustness over
the corresponding subintervals of Table 7.1. These LNARCs yield 30% of the best
FNARC performance curve and a performance improvement of at least 2.12 to 23.8
times better than that of the GNARC design.
Note that the more the desired performance improvement percentage is, the more
number of covering LNARC compensators is needed, consequently leading to an RM-
MAC with more complexity; see Section 7.7 in which different cases of the % FNARC
designs (with higher X%) are chosen that could be used to higher-performance RM-
MAC’s.
Four LNARCs, K1 (s), . . . , K4 (s), are subsequently designed based on X% =
30% of the FNARC, again using mixed-µ synthesis, for each subinterval defined
230
Table 7.1: Best Ap ’s of GNARC and each LNARC for the NMP plant
Compensator Interval PG% Ap
GNARC Z0 = [1.0, 100.0] – 12.5
LNARC#1 Z1 = [4.1, 100.0] 30 298.5
LNARC#2 Z2 = [2.38, 4.10] 30 107.1
LNARC#3 Z3 = [1.34, 2.38] 30 42.6
LNARC#4 Z4 = [1.00, 1.34] 30 26.5
in eq. (7.11); These four LNARCs yield a feedback system with “local” guaran-
tee of stability- and performance-robustness and the best possible local performance
weights (assuming that the rest of the weights are not changed).
Figure 7.8 compares the GNARC design with the four LNARCs by the compen-
sator gain Bode plots.
3
10
GNARC
LNARCs
2
10
#1
1
10
#2
Mag.
0
10
−1 #3
10
−2
10
#4
−3
10
−2 0 2 4
10 10 10 10
Freq. (rad/sec)
Note that, as expected, at low frequencies the LNARCs generate a low-frequency loop-
gain about 2 to 24 times as large compared to the GNARC and this, naturally, leads to
the performance improvements discussed above. It is emphasized that each individual
LNARC closed-loop design has guaranteed performance- and stability-robustness over
its associated parameter subinterval of eq. (7.11). Disturbance-rejection plots (d → y)
will be shown in Section 7.5. Also note from Figure 7.8 that the LNARCs are quite
different at higher frequencies as well.
where Zi∗ = Z ∗ (i, 1) is the nominal NMP zero used in the KF#i, so that the BPMs
are identical at the boundary of each subinterval (model) as shown in Figure 7.9.
The bounds defined in eq. (7.11) were used to design the bank of the four steady-
state discrete-time KFs required for the MMAE implementation. The identification
subsystem of RMMAC will then generate in real-time the four posterior probabilities
Important Remark: The KFs are designed based on the BPM such that the BPMs
agree at the model boundaries of eq. (7.11). Again, note that the nominal parameters
232
−7.5
Model#1
Model#2
Baram Proximity Measure (BPM) Model#3
Model#4
−7.8
−8.1
0 1 2
10 10 10
Z
Figure 7.9: Baram proximity measures (BPMs) based on the nominal NMP zeros
given in eq. (7.12) used in the KF designs.
of the KF models are not necessarily the same as the midpoints of eq. (7.11). The
posterior probabilities will converge to the “nearest probabilistic neighbor” of one of
the “nominal” designs.
By having the LNARCs and the KFs designed, the RMMAC is ready to be constructed
as shown in Figure 7.10. Meanwhile, as mentioned repeatedly, one of the RMMAC
capabilities is that we can predict the potential performance of the RMMAC over the
best GNARC.
First, let us calculate a set of the two most important closed-loop transfer func-
tions with engineering relevance. The first is the closed-loop transfer function Myd (s)
from the plant disturbance d(t) to the plant output y(t), which measures the qual-
ity of the disturbance-rejection. The second Mud (s) is the transfer function from
the plant disturbance d(t) to the plant control u(t), which measures the frequency-
domain characteristics of the demanded control signals. These two transfer functions
233
ξ (t) θ (t)
u (t ) y (t )
Unknown plant
r1 (t ) u1 (t ) u (t )
KF#1 K1(s)
r2 (t ) u2 (t )
KF#2 K2(s)
r3 (t ) u3 (t )
KF#3 K3(s)
r4 (t ) u4 (t )
KF#4 K4(s)
S1 P1 (t )
Residual S2 Posterior P2 (t ) Posterior
covariances Probability P3 (t ) probabilities
S3
Evaluator
S4 (PPE) P4 (t )
Figure 7.10: The four-model RMMAC architecture for the NMP example.
1 1
10 10
0 0
10 10
Mag.
Mag.
−1 −1
10 10
−2 −2
10 10
GNARC GNARC
LNARC#1 LNARC#2
−3 −3
10 10
−3 −2 −1 0 1 2 −3 −2 −1 0 1 2
10 10 10 10 10 10 10 10 10 10 10 10
Freq. (rad/sec) Freq. (rad/sec)
0 0
10 10
Mag.
Mag.
−1 −1
10 10
GNARC GNARC
LNARC#3 LNARC#4
−2 −2
10 10
−3 −2 −1 0 1 2 −3 −2 −1 0 1 2
10 10 10 10 10 10 10 10 10 10 10 10
Freq. (rad/sec) Freq. (rad/sec)
From Figure 7.11 we can also see the characteristics of the GNARC, for different
values of the zero, which has robust stability and performance for the original large
uncertainty upon the zero value, eq. (7.2). Reducing the parameter uncertainty to
the much smaller interval of Model#4 in eq. (7.11) yields a small improvement in
performance as shown in Figure 7.11(d). Thus the GNARC “hedges” for the worst
value of the NMP zero, namely Z = 1, and very little improvement can be expected
through the use of the LNARC design of Model#4. This means, that if indeed the
NMP is in the interval of Model#4, no adaptive scheme can deliver truly superior
performance. The low-frequency zero dominates! The reader should also note the
“push-pull” effect of the NMP zero. Thus, at higher frequencies the LNARC designs
have a somewhat worse disturbance-rejection than the GNARC design. As it is
evident from the remainder of the plots in Figure 7.11, as the NMP moves away from
the jω-axis, the LNARC designs have better and better low-frequency performance in
235
Figure 7.12 evaluates the potential performance improvement of using RMMAC us-
ing stochastic metrics, namely by comparing the RMS errors of the output z(t) and
the control u(t), for different values of Z when the only stochastic input is assumed
to be ξ(t) (i.e. setting θ(t) = 0). These RMS results are readily computed by solv-
ing standard covariance algebraic Lyapunov equations for stochastic LTI systems [56].
Besides, these analytical RMS results must be consistent with those obtained through
time-domain simulation results, at steady-state, as were verified in the simulations.
The graphs of Figure 7.12 vividly suggest that RMMAC has the potential of decreas-
ing the output RMS by a factor of 0.9 to 3.5 while using slightly large control RMS,
by a factor of 1.1 to 1.2, over the GNARC system.
Recall that RMS errors are related to the area of the plots in Figure 7.11. Since
the LNARCs are H∞ compensators, there is no guarantee that the LNARC RMS
errors would always be smaller than that of the GNARC1 .
Figure 7.13, similar to Figure 7.12, evaluates the potential performance improve-
ment of using RMMAC by comparing the RMS errors of the output z(t) and the
control u(t), for different values of Z when plant disturbance ξ(t) and the sensor
1
Only H2 compensators minimize RMS errors, by design.
236
0 0.35
10
GNARC
LNARC 0.3
RMS of u(t)
RMS of z(t)
−1
10 0.25
0.2
−2
10 0.15
0 1 2 0 1 2
10 10 10 10 10 10
Z Z
(a) Output RMS comparisons (b) Control RMS comparisons
Figure 7.12: Predicted potential RMS performance of the RMMAC vs GNARC due
to ξ(t)only.
noise θ(t) are both stochastic inputs to the system. The graphs of Figure 7.13 in-
dicate that RMMAC has the potential of decreasing the output RMS by a factor of
0.9 to 3.5, almost identical to the case when it was due to ξ(t) only; i.e. the sensor
noise does not affect the behavior of the closed-loop transfer function from ξ(t) to
output z(t). Besides, comparing the control RMS, Figure 7.13(b) demonstrates that
the RMMAC will use the same control by exploiting the LNARC#2, #3 and #4.
However, LNARC#1 uses more control than the one in which only the stochastic
input was considered to be ξ(t) as shown in Figure 7.12(a); this is expected since for
smaller NMP zeros the LNARCs do not have good performance improvement (over
the GNARC) due to the NMP zero limitation, but LNARC#1 was shown to have
good performance although it might use larger control than the GNARC. This is due
to the large bandpass characteristics of LNARC#1 as shown in Figure 7.8.
0 0.35
10
GNARC
LNARC 0.3
RMS of u(t)
RMS of z(t)
−1
10 0.25
0.2
−2
10 0.15
0 1 2 0 1 2
10 10 10 10 10 10
Z Z
(a) Output RMS comparisons (b) Control RMS comparisons
Figure 7.13: Predicted potential RMS performance of the RMMAC vs GNARC due
to both ξ(t) and θ(t).
237
7.5.2 Mismatch-Instability
Additional insight into the overall RMMAC performance can be obtained from the
“mismatched-model stability table” shown in Table 7.2. In this table we allow the
uncertain zero to vary over the parameter subintervals of eq. (7.11). We then evaluate
the stability of the resulting feedback system by checking the resulting closed-loop
poles as we did in Chapter 6 to examine the problem of mismatch-instabilities .
Of course, the diagonal entries of Table 7.2 are always stable (S), by construction.
For example, the third row means that when the zero belongs to Model#3, Z ∈
Z3 = [1.34, 2.38], the closed-loop system is always stable if LNARC#3 or #4 is used,
while is always unstable (U) if LNARC#1 is used. If the LNARC#2 is placed in the
feedback we have instability for larger values of NMP Z, but have stability for smaller
values (CU). Thus, in the RMMAC context, as long as P3 (t) → 1 or P4 (t) → 1, the
closed-loop system will be stable. This is due to the fact that the mixed-µ upper
bound inequality in eq. (5.5) is only a sufficient condition for robust stability and,
hence, each LNARC design will actually have a wider robust stability region.
The stability pattern in Table 7.2 can be easily explained on intuitive grounds.
The compensators, say, LNARC#1 or#2 were designed for higher performance and,
hence, have a higher gain that can drive plants with very low-frequency NMP zeros
unstable. On the other hand, the first row of Table 7.2, corresponding to a larger
NMP zero, the closed-loop system remains stable with LNARC#1 or#2, because
they are low gain compensators, designed for the worst condition case of NMP zeros
close to the jω-axis.
Table 7.2: The mismatched-model stability cases for the NMP plant
Model LNARC
# #1 #2 #3 #4
1 S S S S
2 CU S S S
3 U CU S S
4 U U CU S
S : always stable
U : always unstable
CU : conditionally unstable
238
In this example, Table 7.3 summarizes the actual stability margins for the LNARCs.
Of course, performance-robustness is only guaranteed for the subintervals [Zl Zr ] de-
fined in eq. (7.11).
1 1
0.8 0.8
0.6 0.6
P (t)
P2(t)
1
0.4 0.4
0.2 0.2
0 0
0 2 4 6 8 10 0 2 4 6 8 10
time (sec) time (sec)
1 1
0.8 0.8
0.6 0.6
P (t)
P4(t)
3
0.4 0.4
0.2 0.2
0 0
0 2 4 6 8 10 0 2 4 6 8 10
time (sec) time (sec)
Figure 7.14: Case 1A. Transient behavior of RMMAC posterior probabilities with
Z = 4.89.
In this case, we shall show the results of the RMMAC, condensed in Figure 7.10, under
the normal conditions, i.e. all theoretical assumptions are satisfied. The performance
of the RMMAC was evaluated by a very large number of stochastic simulations.
Space limitations allow us to present only typical results. In all cases, the posterior
probabilities exhibited the type of convergence suggested by the theory when all
assumptions are satisfied.
CASE 1A. In this case, the true value of the NMP zero is the nominal of
Model#1, Z = 4.89. Figure 7.14 demonstrates the time-evolution of the four posterior
probabilities. Clearly, P1 (t) → 1. The other three probabilities P2 (t), P3 (t), P4 (t) → 0
very rapidly from their initial values of 0.25.
Figure 7.15 shows a comparison of the GNARC and the RMMAC response under
240
0.4
0.2
0
z(t)
−0.2
RMMAC
−0.4 GNARC
Figure 7.15: Case 1A. Performance characteristics, the output z(t) vs time due to
stochastic inputs. The true value of the NMP zero is Z = 4.89.
CASE 1B. In this case, the true value of the NMP zero is Z = 100.0 in Model#1.
This is the case in which we expect the RMMAC to have truly superior performance
since we are 2 decades above the minimum value Z = 1. Figure 7.16 shows the
time-evolution of the posterior probabilities. Clearly, P1 (t) → 1 while the other three
probabilities decay very rapidly from their initial values of 0.25. In this case, the
nearest probabilistic neighbor to the true system with Z = 100 is Model#4 (with
nominal Z = 4.89 used in the KF#4 design).
Figure 7.17 shows the plant output for the GNARC and the RMMAC; the perfor-
mance improvement of the RMMAC is better and consistent with the expected one,
see Figure 7.11, and better than that of Case 1A, since the NMP zero is larger. Once
more, averaging 10 Monte Carlo runs yields very similar transients (not shown) for
the posterior probabilities and responses.
The control signals (not shown) for both Cases 1A and 1B did not exhibit any un-
usual behavior. The amplitude of the RMMAC controls were about 10% higher than
the corresponding ones for the GNARC, as expected since RMMAC yields improved
disturbance-rejection.
241
1 1
0.8 0.8
0.6 0.6
P (t)
P (t)
1
2
0.4 0.4
0.2 0.2
0 0
0 2 4 6 8 10 0 2 4 6 8 10
time (sec) time (sec)
(a) Model#1 probability, the correct one in this (b) Model#2 probability
case
1 1
0.8 0.8
0.6 0.6
P (t)
P (t)
3
0.4 0.4
0.2 0.2
0 0
0 2 4 6 8 10 0 2 4 6 8 10
time (sec) time (sec)
Figure 7.16: Case 1B. Transient behavior of RMMAC posterior probabilities. The
true value of the NMP zero is Z = 100.0
0.3
0.15
0
z(t)
−0.15
−0.3 RMMAC
GNARC
Figure 7.17: Case 1B. Performance characteristics of the GNARC and the RMMAC,
the output z(t) vs time due to stochastic inputs. The true value of the NMP zero is
Z = 100.0
242
Similar stochastic Monte Carlo simulations (not shown) were made for a variety of
cases for smaller values of the NMP zero. In almost all cases the results were consistent
with the reduced performance implied by Figure 7.11. The actual unmodeled time-
delay in all simulations was selected τ = 0.01 secs at the plant input. We also
tested RMMAC with the maximum possible time-delay, i.e. τ = 0.05 secs. In some
cases, when the values of the actual NMP zero were near the upper boundary of the
intervals of eq. (7.11), the posterior probabilities converged to the next higher model,
apparently a closer probabilistic neighbor than the expected. The reason why the
next higher model was selected in the presence of maximum time-delay is explained
below.
A pure time-delay of τ seconds satisfies the equation y(t) = u(t−τ ). This system is
a linear input/output system and its transfer function is Gτ (s) = e−sτ . Unfortunately,
for simulation purposes, Matlab is not able to perfectly represent a time-delay in this
form, and so we have to use a Pade approximation, which gives a constant gain transfer
function with phase that approximates a time-delay. Without losing generality, using
a 1st order Pade approximation, the time-delay can be approximated as
e−sτ /2 1 − sτ /2
G1τ (s) = sτ /2 ' (7.15)
e 1 + sτ /2
where this function can be computed using the Pade function in Matlab.
Assume that there is a time-delay of τ seconds, which will be inserted in the input
of plant. This augments another NMP zero at Z = 2/τ to the plant. Now, the plant
contains two NMP zeros with the resulting NMP zero larger than the actual NMP zero
of the plant. Hence, in cases with large time-delay inserted, the posterior probabilities
converge to the next model (with higher vales of the NMP zeros). Nevertheless, as
expected from Table 7.2, no instabilities were noted under normal conditions, i.e.
when all assumptions are satisfied.
243
Researchers familiar with MMAC methods are often concerned that the posterior
probabilities may lock-on upon a “bad” model, which will cause long-term instabil-
ity. In our view, this is very unlikely (nothing is absolutely certain in a stochastic
model). If the closed-loop system becomes temporarily unstable, the output and
control signals will grow, thus providing additional excitation to the plant dynamics
and improving the signal-to-noise ratio. As a consequence, the posterior probabilities
should converge to a “good” neighbor. The following numerical example illustrates
the above discussion.
CASE 2A. In this case, the worst possible location for the NMP zero, Z = 1, is
selected. Its nearest probabilistic neighbor is Model#4, with the KF nominal Z = 1.1.
Table 7.2 shows that the combination of Model#4 and LNARC#1 yields an unstable
closed-loop system.
In order to force such an initial unstable mismatch the initial values of the posterior
probabilities are selected as:
rather than the usual initial values of 0.25 used in Cases 1A and 1B . Thus, it is
expected that the initial mismatch will cause instability (hopefully temporarily).
Figure 7.18 shows the transients of the posterior probabilities for a set of 10
different Monte Carlo runs, as well as their numerical average. In all cases,
Figure 7.19 shows the output and control signals, averaged over 10 Monte Carlo
runs in which, after an initial period of brief “instability”, the RMMAC recovers and
returns to its predictable disturbance-rejection. Clearly no sign of long-term insta-
bility is apparent and the self-adjusting property of the RMMAC is demonstrated.
244
1 1
0.75 0.75
P1(t)
P2(t)
0.5 0.5
0.25 0.25
0 0
0 2 4 6 8 10 0 2 4 6 8 10
Time (sec) Time (sec)
1 1
0.75 0.75
P3(t)
P4(t)
0.5 0.5
0.25 0.25
0 0
0 2 4 6 8 10 0 2 4 6 8 10
Time (sec) Time (sec)
Figure 7.18: The behavior of 10 Monte Carlo samples for the posterior probabilities
for Case 2A. The true value of the NMP zero is Z = 1.0. Note that P1 (t) rapidly
decays from its initial value P1 (0) = 0.97 to zero in less than a second. The numerical
average is also shown (solid curves).
Since the NMP zero is at its worst location, Z = 1, performance is not fantastic, as
expected.
P3 (T ) = 0.97, P1 (T ) = P2 (T ) = P4 (T ) = 0.01
at time T = 30 secs for 0.2 secs i.e. the RMMAC system is forced to be unstable at
245
0.4 0.3
0.2
0.15
0
z(t)
u(t)
0
−0.2
−0.15
−0.4
−0.3
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Time (sec) Time (sec)
(a) The average output z(t) (b) The average control u(t)
Figure 7.19: The performance of the RMMAC for Case 2A averaging 10 Monte Carlo
runs. The true value of the NMP zero is Z = 1.0
time t = T ; see Tables 7.2 and 7.3. Thus, it is expected that the forced mismatch
will cause instability, possibly temporarily.
Figure 7.20 shows the transients of the posterior probabilities for a set of 10
different Monte Carlo runs, as well as their numerical average. In all cases, it can
be seen that despite the forced instability for all 30 ≤ t ≤ 30.2, using LNARC#3-
Model#1 mismatched combination, the RMMAC recovers and returns to its correct
predictable superior disturbance-rejection in about 5 secs.
Figure 7.21 shows the output and control signals, averaged over 10 Monte Carlo
runs. It can be seen that after a period of brief “instability”, starting at t = 30 secs,
no sign of long-term instability is apparent and the self-adjusting property of the
RMMAC is demonstrated. It is noted again that since the NMP zero is at its worst
location, Z = 1, performance is not fantastic, as expected.
In the results reported above, and other examples that have not been included in
this section due to space limitations, the overall performance of the RMMAC algo-
rithm was very satisfactory. No behavior inconsistent with the available theory was
observed. However, it is stressed that the results obtained satisfied the probabilistic
assumptions for MMAE convergence; see Chapter 3.
246
1 1
0.75 0.75
P1(t)
P (t)
0.5 0.5
2
0.25 0.25
0 0
0 5 10 15 20 25 30 35 40 45 0 5 10 15 20 25 30 35 40 45
Time (sec) Time (sec)
0.75 0.75
P3(t)
P (t)
0.5 0.5
0.25 4 0.25
0 0
0 5 10 15 20 25 30 35 40 45 0 5 10 15 20 25 30 35 40 45
Time (sec) Time (sec)
Figure 7.20: The behavior of 10 Monte Carlo samples for the posterior probabilities
for Case 2B. Note that P3 (t) rapidly decays from its forced value P3 (0) = 0.97 to zero
in less than a second. The numerical average is also shown (solid curves). The true
value of zero Z = 1.
In this section, the RMMAC is tested when some of the assumptions are intention-
ally mildly violated for the posterior probability convergence. We shall demonstrate,
via examples, that the RMMAC is still capable to deliver excellent performance, even
though a key assumption is mildly violated, by changing the plant disturbance signal
to also include an (unknown to the MMAE) deterministic component, where the sen-
sor noise is sinusoidal (rather than white), and also the uncertain Z is a time-varying
(sinusoidal) zero.
To be more specific, in the remainder of this section it is assumed that the actual
disturbance da (t), driving the plant, is a low-frequency sinusoid superimposed with
the original stochastic disturbance as in eq. (7.4), i.e.
3 4
2
3
1
2
u(t)
z(t)
0
1
−1
−2 0
−3 −1
0 10 20 30 40 50 60 0 10 20 30 40 50 60
time (sec) time (sec)
(a) The average output z(t) (b) The average control u(t)
Figure 7.21: The performance of the RMMAC for Case 2B, averaging 10 Monte Carlo
runs, when we enforce temporary instability at t = 30 secs. The true value of zero
Z = 1.
1 1
0.75 0.75
P1(t)
P2(t)
0.5 0.5
0.25 0.25
0 0
0 2 4 6 8 10 0 2 4 6 8 10
Time (sec) Time (sec)
1 1
0.75 0.75
P3(t)
P4(t)
0.5 0.5
0.25 0.25
0 0
0 2 4 6 8 10 0 2 4 6 8 10
Time (sec) Time (sec)
Figure 7.22: The transient behavior of the posterior probabilities. The plant dis-
turbance includes a deterministic sinusoidal component plus filtered white noise,
eq. (7.17). Z = 10.
magnitude of the control signal u(t) of the RMMAC is about the same as the actual
disturbance d(t) and that it is almost exactly out of phase (180 degree behind) the
actual disturbance that leads to a better disturbance-rejection by using the RMMAC
that is shown in Figure 7.23(b); see the block diagram of the NMP plant in Figure 7.1.
Similar behavior was observed when 10 Monte Carlo simulations were averaged,
and also when the actual location of the NMP zero was changed. So, based upon our
limited experience with this example, it is concluded that the RMMAC is “robust”
even if some of the key assumptions of the MMAE algorithm are mildly violated.
249
3
d (t) GNARC
a
0.4 RMMAC
2 y (t)
open
1 0.2
d(t) & yopen(t)
y(t)
0 0
−1
−0.2
−2
−0.4
−3
0 50 100 150 200 0 50 100 150 200
Time (sec) Time (sec)
(a) The true disturbance and open-loop output (b) GNARC and RMMAC outputs
3
GNARC
RMMAC
2
1
u(t)
−1
−2
−3
0 50 100 150 200
Time (sec)
Figure 7.23: Transient responses of the output and control signals when the plant
disturbance is a sinusoid plus filtered white noise and Z = 10.
Here, we assume that the actual disturbance da (t), driving the plant, is a deterministic
periodic ±5 square-wave disturbance with a period of 33.33 secs superimposed with
the original stochastic disturbance as in eq. (7.4), i.e.
d(t) 2
−2
−4
0 50 100 150 200
Time (sec)
(a) Actual disturbance d(t).
1
P (t)
0.5
1
0
1
P (t)
0.5
2
0
1
P (t)
0.5
3
0
1
P (t)
0.5
4
0
0 10 20
Time (sec)
1.5
0.5
z(t)
−0.5
−1 GNARC
RMMAC
−1.5
0 40 80 120 160 200
Time (sec)
(c) Output response
The actual used disturbance is shown in shown in Figure 7.24(a). Note that the
probabilities converge to the correct Model#1 very fast, i.e. P1 (t) → 1 as shown
in Figure 7.24(b). The RMMAC exhibits an excellent performance improvement
compared to the GNARC as shown in Figure 7.24(c), since Z = 100 is far from the
origin of the s-plane.
Let us test the “robustness to the assumptions” of the RMMAC by using a high-
frequency sinusoid for the measurement noise rather than white noise as the theory
requires. In this case, the disturbance d(t) was a stochastic process as in Section 7.6.1.
However, instead of using white measurement noise, we used in eq. (7.5) a determin-
istic sinusoid measurement error given by:
The KFs in the RMMAC were designed under the assumption that the sensor noise
was white, as in Section 7.4.
The value of the true zero is selected to be Z = 20 in order to clarify the perfor-
mance improvement of the RMMAC over the GNARC. Other values of Z were also
examined in which the RMMAC performed very well (results not shown).
Figure 7.25 shows 10 MC simulation results with the sinusoidal sensor noise as in
eq. (7.19). Figure 7.25(a) shows the four posterior probabilities in which the RMMAC
identifies the correct model (M#1) in about 2 secs. The RMMAC has a better
disturbance-rejection over the GNARC as shown in Figure 7.25(b). Figure 7.25(c)
compares the control signals where the control signal used by the RMMAC is slightly
larger than that of the GNARC; this is consistent with the plots of control RMS that
was shown in Figure 7.13(b) in which the sensor noise input was a pure white noise
including a single frequency of 10 rad/sec as well.
252
0.1
RMMAC
GNARC
0.05
z(t)
1 0
P1(t)
0.5
−0.05
0
0 2 4 6 8 10
−0.1
1 0 20 40 60 80 100
Time (Sec)
P2(t)
0.5
(b) The output z(t)
0
0 2 4 6 8 10 0.2
1
0.15
P3(t)
0.5 0.1
0.05
0
u(t)
0 2 4 6 8 10
0
1
−0.05
P4(t)
0.5 RMMAC
−0.1
GNARC
0 −0.15
0 2 4 6 8 10 0 20 40 60 80 100
Time (sec) Time (sec)
Figure 7.25: The RMMAC evaluation under mild assumption violating with sinusoidal
sensor noise with Z = 20. An average of 10 MC runs are shown.
In all the numerical simulations in this section presented up to now, the uncertain
NMP zero was constrained to remain constant for all time.
As mentioned earlier, the engineering requirement for adaptive control systems is
to deal with slow changes in the plant uncertain parameters. Of course, the presence
of a time-varying non-minimum phase zero violates the plant LTI assumption, and
hence all stationarity and ergodicity assumptions required to prove the posterior
probability convergence results do not hold (see Section 3.5.3). Nevertheless, it is
important to understand, for any adaptive system, its behavior and performance in
the presence of slow NMP zero variations.
In the following, 10 numerical MC simulation results are shown. The uncertain
253
as shown in Figure 7.26(a). Note that this time-varying NMP zero is a sinusoidally
varying zero that can take any value in [1, 100]. The dashed-lines indicate the times
that the zero crosses the boundaries of the four models of Table 7.1. Figure 7.26(b)
shows the dynamic evolution of the four posterior probabilities indicating that the
RMMAC performs very well even when we violate the assumptions by using a de-
terministic measurement noise. Figure 7.26(c) compares the GNARC and RMMAC
performance output responses, which shows that the RMMAC continues to accom-
plish disturbance-rejection very well by using slightly larger control, as shown in
Figure 7.26(d).
It might be attractive to interpret Figure 7.26(b) as demonstrating a “transient”
in the identification process. This may well be true, but one should realize that
the signals generated by the time-varying plant cannot be interpreted using “fixed
model” reasoning. After all changing the NMP zero according to eq. (7.20) implies
an exogenous energy transfer, as a function of time, which is not accounted for in
a “time-invariant parameter” model. This is an important open problem for future
research. Nonetheless, the results of Figure 7.26 (and many others not presented here)
demonstrate the ability of the RMMAC in dealing with slowly-varying uncertain zeros.
In this set of simulation, we want again to evaluate the RMMAC performance when
the NMP zero jumps between the models as shown in Figure 7.27(a). Roughly speak-
ing, we shall see how the RMMAC performs in the presence of large parameter un-
certainty and sudden parameter transition through hypothesis testing in the NMP
system.
The transients of the posterior probabilities are shown in Figure 7.27(b). It is
worthwhile to mention that the RMMAC is always able to identify the true system
quickly, even if the NMP zero has step changes. Figure 7.27(c) compares the out-
puts of the RMMAC with that of the GNARC, in two different Z-axis scales. Note
254
P1(t)
0.5
0
0 200 400 600 800 1000
1
P2(t)
0.5
0
0 200 400 600 800 1000
20 1
P3(t)
0.5
0
Z(t)
P4(t)
2.38
0.5
1.34
1 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time (sec) Time (sec)
(a) Sinusoidal zero, Z(t). Note that Z-axis is log- (b) Posterior probabilities transients
arithmic.
1 0.4
1
GNARC
GNARC
RMMAC
0.5
RMMAC 0.5
0.2
u(t)
z(t)
0 0
0
−0.5 −0.5
−0.2
−1 −1
600 650 700 750 800 850 900 950 1000 0 250 500 750 1000
Time (sec) Time (sec)
that when the NMP zero is small, the RMMAC performance is not satisfactory, as
expected. However, for larger values of Z, the RMMAC outperforms the GNARC.
The two experiments above, one with the sinusoidal time-varying and the other
one with step changes in the NMP zero, are illustrative. These results indicate that
the RMMAC can quickly identify the true (unknown) plant and thus, it can be a
good adaptive design for control applications.
P (t)
0.5
1
50 0
1
40
P (t)
0.5
2
0
30
Z(t)
P (t)
20 0.5
3
0
10 1
P (t)
0.5
4
0 0
0 100 200 300 400 0 100 200 300 400
time (sec) time (sec)
(a) Step changes in the NMP zero, Z(t). (b) Posterior probabilities transients
1 0.4
0.5
0.2
0
z(t)
0
−0.5
−0.2
−1
RMMAC RMMAC
GNARC GNARC
−1.5 −0.4
0 50 100 150 200 200 250 300 350 400
time (sec)
(c) Output response, z(t)
In this section, it is shown that, in some extreme cases, if the theoretical assump-
tions are severely violated, the RMMAC does not work well or goes unstable. As
in Sections 6.3.5.3 and 6.4.4.3 we want to stress that if we violate the theoretical
assumptions in a severe way, the RMMAC can result in a poor performance and/or
instability. We include these “negative results” to point out that great care must be
exercised in the KF designs for the RMMAC.
As we suggest in Chapter 10, by monitoring the residuals on-line one can devise
“fail-safe” logic that results in a “safety net” and switch to the GNARC until the
residuals return to a “normal” range and then resume the RMMAC operation.
256
This case illustrates another very severe mismatch-instability result, similar to Case 2B
in Section 7.6. Here, the RMMAC is forced to use an unstable mismatch at time
t = 30 secs lasting for 10 secs; see Tables 7.2 and 7.3. The true value of Z = 1 is
used. We want to see if this forced long-term mismatch, for all 30 ≤ t ≤ 40, causes
instability in the closed-loop or not. Simulation results for a single Monte Carlo run
are shown in Figure 7.28.
To force this long-term mismatch instability, the values of the probability vector
are selected to be
The transients of the posterior probabilities are shown in Figure 7.28(a)–(d) which
shows that a probability confusion starting at t = 30 secs is apparent and that the
RMMAC can not recover after t = 40 secs. Figure 7.28(e) shows the output of the
RMMAC and that the closed-loop system is unstable with this long forced instability.
The four residual signals shown in Figure 7.28(f) demonstrate that the residual signals
are vastly increased after t = 40 secs and stabilization is not achieved. It turns out
that the MMAE subsystem of the RMMAC will fail in this case. This happens when
all the residuals are very large compared to their computed covariances; recall the
PPE equation; see eq. (3.34).
Of course, the probability of a stochastic event that causes the P3 (t) = 1 for
10 secs (1000 measurements) is truly infinitesimal!
A similar experiment where the forced instability was only for 5 secs did not cause
the RMMAC to be unstable (results not shown here).
In the sequel, we will see that other severe violations of the theory can lead to
probability confusion and instability.
Up to the present, we assumed that the intensity of the plant white noise was known
and constant in both designing the KFs and the actual simulations. It is of interest
257
1 1
P2(t)
0.5 0.5
0.25 0.25
0 0
0 5 10 15 20 25 30 35 40 45 0 5 10 15 20 25 30 35 40 45
Time (sec) Time (sec)
0.75 0.75
P3(t)
P4(t)
0.5 0.5
0.25 0.25
0 0
0 5 10 15 20 25 30 35 40 45 0 5 10 15 20 25 30 35 40 45
Time (sec) Time (sec)
0.2
5
0.1
z(t)
r(t)
0 0
−0.1
−5
0.2
−10 −0.3
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time (sec) Time (sec)
(e) The output z(t) (f) The four residual signals r(t)
Figure 7.28: The RMMAC response following long-term forced instability (30 ≤ t ≤
40). The RMMAC becomes unstable. The true Z = 1 in Model#4 which is correctly
identified before the forced unstable combination. Only one MC run is shown.
to see what happens if the KFs are designed based on a specific white plant noise
intensity, but in the simulations the RMMAC is faced with a disturbance generated
by a drastically different white noise intensity. In this section, two different cases
are presented. In both cases, the KFs are designed based on the known and fixed
white noise plant intensity Ξ = 1 but we use either smaller or larger intensities in the
simulations.
Remark: Since the plant noise intensity is not changed for designing the KFs the
BPM curves are the same as in Figure 7.9. Also the Kalman Filters are designed
based on the nominal points as given in eq. (7.12).
CASE 3A. In this case, the intensity Ξ = 100 is used in the simulations, i.e. the
random disturbance signal is 10 times (in standard deviation) larger the one used in
258
the KFs design. The true value of the NMP zero is selected to be Z = 50. In fact we
look for the cases in which the posterior probabilities do not converge and/or oscillate
leading the RMMAC to, possibly, an unstable closed-loop. For smaller values of NMP
zero we do not observe any oscillation in the probabilities that creates instability. For
this reason, we decide to select the large value for NMP zero, Z = 50, used in the
simulation.
The simulation results are shown in Figure 7.29. The transients of the posterior
probabilities are shown in Figure 7.29(a)–(d). It can be seen that there is a model
selection confusion among the posterior probabilities. Figure 7.29(e) compares the
GNARC and RMMAC outputs. In the RMMAC response, temporary instabilities
are observed at those times that the probabilities oscillate. Note that the RMMAC
has poor performance compared to the GNARC. However, it does not go unstable.
The reason is that all LNARCs stabilize the plant with the NMP zero at Z = 50,
according to Table 7.2. Figure 7.29(f) shows the control signal used by RMMAC.
Figure 7.29(g) shows the residual signals. Note that the residuals are large in some
time-intervals and in the same intervals the probabilities are also oscillating, and are
inconsistent with the residual covariances used in the PPE calculations.
It is again emphasized that we looked for all those cases that the RMMAC might
fail; for smaller values of Z, the RMMAC works very well due to the fact that
the LNARCs are low gain compensators, designed for the worst case of NMP ze-
ros (smaller zeros).
CASE 3B. In this case, again the design plant disturbance white noise intensity
is unity (Ξ = 1) but in the simulations Ξ = 0.01 is used. The true NMP zero is
selected to be Z = 1.50; for larger values of Z, the RMMAC works very well. Similar
to Case 3A above, we search for those cases in which the posterior probabilities does
not converge correctly and/or oscillate leading to, possibly, an unstable closed-loop
system. Again, for larger values of NMP zero we did not see any oscillation in the
probabilities nor instability.
The simulation results are shown in Figure 7.30. The transients of the posterior
259
1 1
0.75 0.75
P1(t)
P2(t)
0.5 0.5
0.25 0.25
0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time (sec) Time (sec)
1 1
0.75 0.75
P3(t)
P4(t)
0.5 0.5
0.25 0.25
0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time (sec) Time (sec)
10
4
2 0
0 −10
−2 −20
u(t)
z(t)
−4
−30
−6
−40
−8
RMMAC −50
−10 GNARC
−12 −60
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
time (sec) Time (sec)
(e) The RMMAC output z(t) (f) The RMMAC control u(t)
0.3
0.2
0.1
r(t)
−0.1
−0.2
−0.3
0 10 20 30 40 50 60 70 80 90 100
Time (sec)
Figure 7.29: Case 3A. The RMMAC evaluation under severe assumption violation.
The plant disturbance intensity is unit (Ξ = 1) in designing the KFs, but in the
simulations a very large disturbance, with Ξact = 100, drives the NMP plant. The
value of the NMP zero is Z = 50. Only one MC run is shown. Note the poor
performance of the RMMAC.
260
probabilities are shown in Figure 7.30(a). It can be seen that there is a model selec-
tion confusion among the posterior probabilities. Figure 7.30(b) shows the RMMAC
output. Again, temporary instabilities are observed but the closed-loop system is not
completely unstable. Note that the compensator LNARC#1 was designed for higher
performance and, hence, has a high gain that can drive plants with very low-frequency
NMP zeros unstable.
For NMP zeros larger than 1.72, all LNARCs can stabilize the NMP plant; hence
we did not observe poor behavior of the RMMAC for NMP values in [1.72, 100]; refer
to Table 7.3.
Figure 7.30(e) shows the control signal used by RMMAC. The residual signals
(not shown) are not very large compared to their computed covariances.
Remark: In this case, persistent excitation is not large enough. Perhaps this is
why the MMAE, as the identification subsystem of the RMMAC, can not perform its
job well.
We did not implement the RMMAC/XI architecture for Cases 3A and 3B. Its use
would probably correct the noted problems.
P1(t)
1 1
0.5 0.5
z(t)
0
0 100 200 300 400 500
1
−0.5
P2(t)
0.5
−1
0 100 200 300 400 500
0 Time (sec)
0 100 200 300 400 500
(b) The RMMAC output z(t)
1
0.8
P3(t)
0.5
0.6
0.4
0
0 100 200 300 400 500 0.2
u(t)
1 0
−0.2
P4(t)
0.5 −0.4
−0.6
0
0 100 200 300 400 500 −0.8
0 100 200 300 400 500
Time (sec) Time (sec)
Figure 7.30: The RMMAC performance evaluation under severe assumption violation
where the designing plant disturbance intensity is unit (Ξ = 1) but a very small,
Ξact = 0.01, is used in the simulations. The value of the NMP zero is Z = 1.50 and
only one MC run is shown.
3
10
2
Ap
10
Best GNARC
FNARC
50% of FNARC
Performace levels
1
10
0 1 2
10 10 10
Z
Figure 7.31: Best performance gains for each of the six LNARCs by requiring X% =
50% of the best FNARC performance curve.
Similar to Section 7.7.1, the plots of the maximized performance parameter Ap for
the FNARC, the GNARC, and also the LNARC designs, based on X%=60% of the
FNARC, are shown in Figure 7.32. Except for the LNARCs, the GNARC and the
FNARC are identical to the best designs in Section 7.3 and Section 7.7.1.
By comparing Figure 7.32 with Figure 7.7 and Figure 7.31 again we can see that
there is a potential performance improvement by using a higher X% of the FNARC,
as expected. However, we have to pay the price by using eight LNARCs.
3
10
Best GNARC
2 FNARC
p
10
A
60% of FNARC
Performace levels
1
10
0 1 2
10 10 10
Z
Figure 7.32: Best performance gains for each of the eight LNARCs by requiring
X% = 60% of the best FNARC performance curve.
The covering subintervals for LNARCs with X% = 60% are summarized in Ta-
ble 7.5. The interval of NMP zero uncertainty is divided into eight subintervals.
Again, there are more parameter intervals near the worst case, Z = 1, required for
robust designs, due to the inherent NMP zero limitation. Table 7.5 indicates that by
264
using a higher X% the reduction in parameter uncertainty allows even larger perfor-
mance parameters Ap for designing the LNARCs. These LNARCs yield 60% of the
best FNARC performance curve and the improvement in performance at least 2.7 to
47.62 times better than that of the GNARC design; in fact this new designed RM-
MAC (with X%=60%) demonstrates 1.27 to 2 times better performance enhancement
in comparison with the last designed RMMAC (with X%=30%); compare Tables 7.1
and 7.5. However, the price is again increased complexity of the RMMAC with more
LNARCs and KFs.
Here, we discuss another approach to design the RMMAC for the NMP plant. In this
approach, the designer essentially considers the adaptive complexity of the multiple-
model system, perhaps guided by the previous results.
Assume that the initial NMP zero uncertainty is given in [Zl , Zu ]. Knowing that
small NMP zeros limit the performance, as well as that the RMMAC hedges at the
small NMP zeros by employing more models, in the Brute-Force method a single
model associated with the smaller values of the zeros, say [Zl , Zm ], is arbitrarily
selected. Recall that the RMMAC will yield high-performance disturbance-rejection
for large zeros. For the reminder of the NMP zero uncertainty, [Zm , Zu ], an X% is
selected following the step-by-step algorithm of the % FNARC method to design the
LNARCs (similar to the previous designs). It is obvious that the total number of
265
models will be the number of models found based on the % FNARC method plus one
that was chosen to cover smaller NMP zero values in [Zl , Zm ].
Here we show a representative Brute-Force approach. One LNARC was first
design to cover the NMP zeros in [1.0, 10.0]; a good reason for this selection can be
seen in Figure 7.6. For the values of Z = [10, 100] we shall desing other LNARCs
based on X%=50% of the FNARC. However, only a single LNARC could yield almost
X%=100% of the FNARC, for all values of zero in [10, 100. Thus, we would need only
two models in this case.
Table 7.6 summarizes the results of both LNARC subintervals. Recall that the
LNARC#1 (for all Z ∈ [10, 100]) yields almost 100% of the best FNARC performance
in [10.0, 100.0]. Also, LNARC#2 (for all Z ∈ [1, 10]) demonstrates the improvement
in performance about 2.3 times better than that of the GNARC design, for the values
of NMP zero in [1.0, 10.0].
3
10
LNARC#1
2
p
10
A
LNARC#2
GNARC
1
10
0 1 2
10 10 10
(II) Z
(I)
Figure 7.33: LNARCs for the NMP plant by using a Brute-Force approach.
Table 7.6: Best Ap ’s of LNARCs for the NMP plant based on a Brute-Force approach
LNARC# Interval Ap
1 Z1 = [10.0, 100.] 995.3
2 Z2 = [1.0, 10.0] 28.8
Assuming that the NMP zero is a random variable with uniform density, it is clear
to see that the RMMAC (in fact LNARC#1) will present almost the best FNARC
performance at about 90% of the location of NMP zeros; this is a very encouraging
266
result. For the rest 10% of uncertainty interval, i.e. [1.0, 10.0], the RMMAC still
performs a better job in comparison to the GNARC, even though it is not significant.
We did not explicitly design, evaluate, and simulate the RMMACs resulting from
the material in Section 7.6.1, 7.6.2, and 7.6.3.
7.8 Conclusions
In this chapter, we examined a SISO non-minimum phase (NMP) plant, with one
uncertain real-valued NMP zero, and we evaluated the performance of a four-model
RMMAC vs. GNARC. When the uncertain NMP zero is close to the jω-axis, the RM-
MAC can only deliver a very slight improvement in disturbance-rejection compared
to the GNARC (which hedged for the smallest possible value of the NMP zero). This
is due to the inherent performance-limitations associated with low-frequency NMP
zeros, see e.g. [108, 125], independent of the design methodology. For larger values of
the NMP zero the RMMAC provided significant improvement in the performance, as
expected and verified by the simulations.
Even though the example considered in the chapter is academic, it provides valu-
able insight into the RMMAC method. The following points are emphasized:
A) Plants with low-frequency NMP zeros are hard to even stabilize, and adaptive
control (or any other control) cannot produce superior performance. However, if the
uncertainty range of the NMP zero is large, then RMMAC can provide performance
improvements over non-adaptive robust designs, when the actual value of the NMP
zero is large.
B) The definition of the number, nominal, and parameter subintervals, as in
eq. (7.11), are very important. For the MMAE, convergence to a nearest proba-
bilistic neighbor does indeed occur, but this must be done by exploiting the Baram
Proximity Measure (BPM).
C) The local compensators (LNARCs) were designed with the mixed-µ method,
which is only a sufficient condition for robust-stability and robust-performance. It
is noted that, at least, stability is maintained on both sides of the corresponding
subinterval of eq. (7.11). So RMMAC appears to be even more robust because of this
267
fact.
D) It is important to predict performance improvements (if any) by the RMMAC
by making off-line comparisons of the type shown in Figure 7.11, computing and com-
paring RMS values and constructing stability tables of the type shown in Table 7.2.
In essence, the designer can predict the precise potential benefits of the sophisticated
RMMAC design without doing a single stochastic simulation.
E) It is essential that RMMAC based designs be verified via extensive simulations
that reflect the true operational environment which will inevitably violate some of
the assumptions, as in Section 7.6.2.
With respect to item B, the number of models exploited in any RMMAC design
should be as small as possible, because the dynamic and numerical complexity of
the RMMAC algorithm increases as the number of models are increased. Because
of the insight obtained by the above results, a high-frequency design (HFD) was
also tested by increased the performance corner frequency as well as disturbance
power frequency to α = 20 rad/sec in eqs. (7.4) and (7.8). As we would expect, the
HFD RMMAC design needs more number of models associated with the subintervals
in comparison to the low-frequency design (LFD) considered in this chapter. Also
significant improvement was obtained in disturbance-rejection compared at higher
values of NMP zero, as expected. The results are not shown due to space limitations.
We also discussed the “Brute-Force” approach in which we associate a model for the
smaller values of the NMP zero. It showed that it could be a better solution for the
NMP cases.
The RMMAC performance conclusions reported here and also in Chapter 6 fur-
ther validate earlier simulations by the authors. In [23] we analyzed, using a five-
model RMMAC, an academic SISO non-minimum phase (NMP) plant with a sin-
gle uncertain right-half plane zero which also posed fundamental limits to superior
disturbance-rejection; see, e.g., [125]. In that example, the GNARC also “hedged” for
the minimum value of the NMP zero. If the true zero was near its minimum value,
the RMMAC performance was essentially the same as for the GNARC (no magic
properties). The true value of the RMMAC shows when the true NMP zero is large;
then the RMMAC produces truly superior performance vis-à-vis the GNARC.
268
Chapter 8
8.1 Introduction
In the chapter, we discuss the multivariable (MIMO) version of the RMMAC design
methodology. In the MIMO case, in general, we have:
. several control variables
. several performance variables
. several plant disturbances
. several noisy measurements
. several unmodeled dynamics
. several uncertain real parameters
The multivariable version of the RMMAC presents numerous pragmatic difficulties
in defining the RMMAC architecture in a systematic step-by-step manner, unlike the
SISO version described in Chapter 5 and fully illustrated and evaluated in Chapters 6
and 7. Indeed the modeling and computational difficulties become apparent when we
consider the two-input two-output (TITO) case, with two real uncertain parameters.
Therefore, in this chapter we shall use the TITO case to illustrate the increased design
complexity of the RMMAC. In this manner, we hope to provide insight for the MIMO
case, especially when we deal with more than two uncertain real parameters.
We remark that a specific TITO RMMAC design will be fully discussed and
269
270
p ∈ Ω ⊂ Rm ; Ω hyperparallelpiped (8.2)
The plant is also assumed to have unmodeled dynamics and (unmeasured) stochas-
tic disturbance inputs ξ(t) and noisy sensor measurements y(t).
Using the mixed-µ synthesis, the robust compensators GNARC and FNARC can
be designed for the large parameter uncertainty of eq. (8.2). Similar to the SISO
case, to carry out the mixed-µ design, we must select frequency-dependent weight
matrices upon the disturbances, sensor noises, outputs and control signals.
In all MIMO designs, we fix the same frequency weight matrices for the sensor
noises and the control signals. We also assume that
α
Wd (s) = I2×2 (8.4)
s+α
where Wp1 (s) = γAp Wp0 (s) (γ ∈ R+ ) and Wp2 (s) = Ap Wp0 (s) and
" #
γ 0
Wp (s) = Ap Wp0 (s) (8.7)
0 1
where Ap is called the performance parameter and Wp0 (s) is a stable, rational
transfer, say a simple first-order weight of the following form
α
Wp0 (s) = (8.8)
s+α
which reflects our specification for good disturbance-rejection for the frequency range
ω ≤ α where the disturbances have most of their power. Note that in eq. (8.7) the
first performance output, z1 (t), could be selected to have γ times smaller (if γ < 1)
or larger (if γ > 1) required performance than that of z2 (t); i.e. selecting γ is directly
related to the desired control accuracy of z1 (t) vs z2 (t).
All unmodeled dynamics must be bounded by their maximum singular values and
properly introduced in the mixed-µ synthesis.
With the above frequency weight matrices, we discuss the robust compensators
designs in the sequel. Note that for designing all these compensators, we will use the
same frequency-dependent weight matrices, except the performance parameter Ap in
the performance weight matrix in eqs.(8.5) and (8.7).
272
pmin
k ≤ pk ≤ pmax
k ; k = 1, 2, . . . , m (8.9)
p1 ∈ Ω1 = [pmin max
1 , p1 ] (8.10)
p2 ∈ Ω2 = [pmin max
2 , p2 ]
In order to have the best possible guaranteed performance using the GNARC,
designed for the uncertain bound of eq. (8.9) or eq.(8.10), the performance parameter
Ap in eq.(8.5) or eq. (8.7) is increased until the sufficient condition of the mixed-µ
bound for both stability- and performance-robustness, so that
Then the resulting GNARC compensator, from the D,G-K iteration, denoted
by K(s), indeed yields a nonadaptive feedback system (GNARC) with guaranteed
stability-robustness and the best possible performance-robustness. If the demanded
performance is too high, while the real parameter errors are too large to allow such
performance guarantees, it may not be possible to design the GNARC. If the com-
pensator is judged to be unable to provide the desired performance, the designer
must then decrease the performance parameter, Ap , i.e. be willing to accept inferior
performance, until the sufficient condition of (8.11) is satisfied.
In particular, if the bounds on the real parameter uncertainties are large, then the
GNARC may yield a robust LTI design with inferior performances. This would be a
signal for using an adaptive control system, and the RMMAC in particular.
The FNARC are computed for the entire uncertain hyperparallelpiped with the
maximized performance parameter Ap . This is the best possible performance achiev-
able if the uncertain parameters were known exactly. The FNARC will then define a
hyper surface over the hyperparallelpiped Ω.
To determine the FNARC we use a dense grid of parameters in eq. (8.9) or
eq.(8.10) and determine the associated robust compensator for each grid point using
exactly the same bounds on unmodeled dynamics and frequency weights employed
in the GNARC design. Thus, we can make fair and meaningful comparisons. For
each grid point we use the complex-µ design methodology and Matlab software
[3, 124], because both the performance weights and bounds on unmodeled dynamics
are complex-valued and there are no real parameter uncertainties. For each grid
point we again maximize the performance-parameter Ap in eq. (8.9) or eq.(8.10)
until the complex-µ upper-bound, µub (ω) is just below unity for all frequencies, say
0.995 ≤ µub (ω) ≤ 1; ∀ω, to be consistent with the GNARC upper-bound employed.
The maximum Ap ’s of the FNARC (AFp ) are illustrated in Figure 8.1 for the
case with 2-D real parameter uncertain. The FNARC is simply a surface over the
rectangle.
Ap
p
2
p
1
So far, the procedure for designing the GNARC and the FNARC mimics the SISO
ideas presented in Chapter 5. By comparing the fixed Ap for the GNARC with the
FNARC surface we can see what is the potential performance improvement for using
a complex adaptive system (recall that the FNARC requires an infinite number of
models in the RMMAC architecture).
For the 2-D uncertain parameter case it easy to visualize the FNARC as illustrated
in Figure 8.1. It is obvious that such a “visual” inspection cannot be done for the
m-D uncertain parameter case; one would have to compare numerical values or use
projections of the FNARC hypersurface.
Γ
1
X% of FNARC
p
A
intersection
p2 p
1
above process yields the intersection of surfaces (the equivalent of Γ1 , Γ2 , . . . are also
surfaces).
Unfortunately, the intersection of these surfaces does not occur along rectangular
parameter subsets as shown in Figures 8.2 and 8.3. However, the mixed-µ software
requires rectangular constraints for the uncertain parameter sets. Thus, we can not
generalize the systematic SISO procedure of Chapter 5 deriving the required number
of models by demanding an adaptive performance of a chosen value of X% of the
FNARC.
p2
p
1
Figure 8.3: Projection of the intersection of the Γ1 surface with X% of the FNARC
surface of Figure 8.2.
276
Thus, we can not use any “single direct” method to find the “model” parameter
boundaries to be used in the mixed-µ design and to be able to design the LNARCs,
even in the 2-D case. Obviously, the same difficulty exist in the m-D parameter case
to a greater extent.
Thus, the model definition must be based on ad-hoc methods, perhaps guided by
the FNARC surface characteristics.
Continuing with the 2-D case, here we consider the X% requirement along the top edge
(p2 = pmax
2 ) and proceed as in the SISO case of Chapter 5. This process results in the
N “vertical” models. This model subdivision, in this case, would be vertical uncertain
rectangles similar to those shown in Figure 8.4 where the 2-D uncertainty space is
divided vertically; i.e. the uncertainty of p2 is kept to be as its global uncertainty
bound for all local models while the uncertainty in p1 is split for extracting the local
models.
As a result, the entire initial uncertainty set of eq. (8.10) is covered by N “ver-
tical” models requiring N local compensators. We shall refer to this as the Vertical
LNARCs (VLNARCs). Clearly, the reduction in parameter uncertainty will allow
larger performance gains for designing the LNARCs, resulting into guaranteed both
stability- and performance-robustness over the subsets of Figure 8.4.
277
There is another heuristic way to obtain the LNARCs. In this method i.e. we use the
X% requirement along the right edge (p1 = pmax
1 ) and proceed as in the SISO case
of Chapter 5. Loosely speaking, it means that we keep p1 completely uncertain and
we split the initial uncertain rectangle horizontally. The results would be horizontal
uncertain rectangles as illustrated in Figure 8.5. We shall refer to this as the Horizonal
LNARCs (HLNARCs).
For the same X%, we would have NV required models, for the VLNARCs, and
NH models, for the HLNARCs. These model numbers may be very different since in
fact the FNARC surface may have its maximum slope along the p1 -axis or p2 -axis.
Thus, for the same X%, the LNARCs with more models will yield higher potential
adaptive performance, but the increased complexity may not be acceptable for the
designer who may select (between VLNARCs and HLNARCs) the one that has the
least number of models.
The above procedure will be further clarified in the example of Chapter 9. How-
ever, we would like to mention that by studying the performance implications of using
the different models defined in Figures 8.4 and 8.5 one may derive “hybrid” versions
as shown in Figure 8.6.
278
Figure 8.6: A “hybrid” definition of models for the 2-D RMMAC architecture.
In Section 5.3 we discussed the Kalman filter (KF) design procedure for the SISO
case. The proper design of each KF in the RMMAC architecture is crucial in order
to satisfy the theoretical assumptions [44, 45, 52] which will imply that the PPE will
lead to the correct model identification.
In this section, we discuss the design of the Kalman filters (KFs) of the RMMAC
methodology in the MIMO case. As well shall see in the sequel, the extension of the
SISO case provided in Section 5.3 gets more complicated when we have two, or more,
uncertain real parameters.
The procedures described in Section 8.2.3 led naturally to the number of models N
adopted, together with the definition of their boundaries. This effectively partitions
the space of the plant transfer functions into N regions Ωi ; i = 1, . . . , N . Each
Ωi defines the Model#i. Note that these regions are naturally defined in the finite
dimensional space of plant parameters. It now remains to identify a representative p∗i
for each region Ωi that will become a nominal model in the discrete-time steady-state
Kalman filters (KFs) design procedure.
Considerable care must be taken to choose the representative models. Ideally,
if they are well chosen, and during the real-time operation, all plant models p in an
arbitrary region Ωi should be identified as the corresponding representative p∗i . Stated
intuitively, each representative p∗i specifies the class of equivalence of all plant models
279
where L(p, p∗i denotes the Baram Proximity Measure (BPM) between p and p∗i ;
see Section 3.5 for more details.
However, finding representative p∗i ’s so as to satisfy the above conditions is not an
easy task in case of two or more uncertain parameters cases.
The following strategy proved to be efficient in practice.
(i) Grid the space of allowable plant parameter variations. For each region Ωi ,
this yields a finite number Nji of plants Dij ; i = 1, 2, . . . , N , j = 1, 2, . . . , Nij .
(ii) Define for each p∗i ; i = 1, 2, . . . , N
that is, S(p∗i ) is the number of plants out of the total number Nji of plants in
Ωi for which the BPM from those plants to p∗i is less than then the BPM from any
other p∗i ; j 6= i. Intuitively, S(p∗i ) is the number of plant models that are identified
(represented by) p∗i during the identification procedure.
The objective is then to minimize the gap between S(p∗i ) and Nji . Stated mathe-
matically,
N
X ¯ j ¯
min f = min ¯N − S(p∗i )¯2 (8.15)
∗ i
p∗ pi
j=1
by proper choice of p∗i ; i = 1, 2, . . . , N . In our work, the above cost functional was
minimized using the genetic algorithm in Matlab toolbox. In the ideal case, f = 0
and S(p∗i ) equals Nji exactly, that is, there is exact matching between the “control”
and “identification” regions. There are of course no guarantees that this can be
achieved. In any case, minimizing the cost function above will yield to the “best
possible” match between controls (LNARCs) and identification (MMAE) regions. In
our experience, minimizing the quantity in eq. 8.15 is a very time-consuming task.
Also, the minimization process does not necessarily become zero (ideal identification)
and S(p∗i ) may be much different than Nji .
280
We found that the gap can often be decreased if, after finding the best model
representatives by running the above algorithm and using each p∗i to define the cor-
responding Kalman filter model, “fake plant white noise” is added to those models.
In fact, the need for such “fake plant white noise” is to compensate for the (large)
unknown parameters within each KF so that the KFs pay more attention to the
information contained in the residual signal.
Figure 8.7 shows the BPMs L1 and L2 for a 2-model MMAE with a 2-D uncertain
parameter p = [ p1, p2 ]. For all pA ∈ Ω1 , i.e. when the true parameter belongs to
Model#1, L (pA , p∗1 ) < L ( pA , p∗2 ) and the posterior probabilities P1 (t) → 1, P2 (t) →
0, so that convergence to Model#1 (defined by KF#1) occurs. Similarly, when pA ∈
Ω2 and L (pA , p∗2 ) < L ( pA , p∗1 ), the posterior probabilities P2 (t) → 1, P1 (t) → 0 and
Model#2 (defined by KF#2) will be selected as the closest neighbor to the true plant,
from a probabilistic point of view. The difficulty is to design the KFs so that the
BPM surfaces intersect at all boundaries of adjacent models.
Chapter 9 provides more details of a TITO case study with an example with 2-D
real parameter uncertainty.
BPM#2
BPM#1
BPM
p
1
Ω p2 Ω
2 1
– For each value of the m-D real parameter, the performance parameter, Ap , is
again maximized until µ upper-bound ∼ = 0.995.
Compare the (maximized) Ap hypersurfaces from the GNARC and the FNARC
designs vs p.
• STEP 5. [Selection of X% ]
Obtain the minimum number, N, of required models and their uncertainty range
in both “horizontal” and “vertical” divisions
– Select the appropriate division which has the least number of models
– The larger X%, the more models are required (in either case)
– We use the same design performance weights and the same frequency weight
matrices as in the GNARC, so that the performance comparisons are “fair”.
– Each LNARC will guarantee a “local” performance that exceeds that of the
GNARC.
– From Step 6 we know the (smaller) parameter subset associated with each
model indexed by k ; k = 1, . . . , N . For each KF we must use a “nominal”
parameter value, p∗k .
– The desired p∗k ’s are obtained by matching values (to the degree possible) of
the BPM between adjacent parameter models; see Figure 8.7.
– The BPMs can be calculated for a fine grids of the 2D uncertain parameter;
they are surfaces
– Matching the surfaces of the BPMs at the boundaries can be done by using
optimization algorithms, e.g the Genetic algorithm and direct search toolbox of
Matlab, though this is a very time-consuming process.
– We may also need to employ “fake plant white noise”, in the presence of large
parameter uncertainties, in order to match the surfaces at the boundaries.
Evaluate and compare the detailed expected performance for each LNARC with
that of the GNARC, using:
– Time-transient responses
9.1 Introduction
285
286
in the RMMAC architecture. Once the number of models has been fixed, and their
boundaries defined, we use the mixed-µ synthesis to determine the TITO “Local
Non-Adaptive Robust Compensators (LNARC)” used to implement the RMMAC. In
Section 9.5 we discuss how we can predict the asymptotic performance of the RM-
MAC either in the frequency domain (using suitable sensitivity singular value plots)
or in a stochastic setting (using RMS values). In Section 9.6 we design the bank of
KFs required in the RMMAC architecture by optimizing the BPM. This is a key step
to ensure that the identification subsystem converges to the ”nearest probabilistic
neighbor.” In Section 9.7 we present a representative sample of stochastic simula-
tions, for both constant and slowly-varying unknown parameters, which demonstrate
the superior performance of the TITO RMMAC. In Section 9.8 our conclusions are
summarized.
e− sτ
e−− ssττ
Figure 9.1: The TITO three-cart MSD system. The spring k1 and the mass m3 are
uncertain.
287
y1 (t) and y2 (t), of the positions x2 (t) and x3 (t), of masses m2 and m3 , respectively.
The control forces u1 (t) and u2 (t) act upon the known mass m1 and the uncertain
mass m3 , respectively.
The disturbance forces, d1 (t) and d2 (t), are independent stationary stochastic
processes generated by driving a low-pass filter,
α
Wd (s) = I2×2 (9.1)
s+α
with continuous-time white noise vector ξ(t), with zero mean and unit intensity
matrix, as follows:
d(s) = Wd (s) ξ(s) (9.2)
The overall state-space representation includes the disturbance dynamics via two
augmented state variables d1 (t) and d2 (t) as follows.
ẋ(t) = Ax(t) + Bu(t) + Lξ(t)
y(t) = C2 x(t) + θ(t) (9.3)
z(t) = C x(t)
1
where
h iT
x(t)= x1 (t) x2 (t) x3 (t) v1 (t) v2 (t) v3 (t) d1 (t) d2 (t) ,
h iT
ξ(t) = ξ1 (t) ξ2 (t) ,
h iT
θ(t) = θ1 (t) θ2 (t) ,
h iT (9.4)
u(t) = u1 (t) u2 (t) ,
h iT
y(t) = y1 (t) y2 (t) , and
h iT
z(t) = z1 (t) z2 (t) .
and
− k1m+k 3 k1
m1
k3
m1
1
A21 =
k1
m2
− k1 +km22+k4 k4
m2
,
(9.5)
k3 k4 k3 +k4
m3 m3
− m3
− b1m+b1 3 b1
m1
b3
m1
A22 = b1 − b1 +bm22+b4 b4 , (9.6)
m2 m2
b3 b4
m3 m3
− b3m+b3 4
03×2
1
0 0
m1 6×2
B=
0 0 ,L = α 0 ,
0 1 0 α (9.7)
m3
02×2
" #
0 1 0
C1 = 02×5 , C2 = C1 .
0 0 1
where ki are the spring constants, bi are the damping coefficients, mi , xi , and vi are
the mass, position, and velocity of mass i = 1,2,3 and uj ; j = 1, 2 are the applied
control forces.
The fixed physical parameters in eq. (9.3) are as follows.
m1 = m2 = 1
k2 = 0.15, k3 = 1, k4 = 0.2
(9.8)
b1 = b2 = b3 = b4 = 0.01
α=1
The upper and lower bound for the uncertain parameters for the spring constant
k1 and the mass m3 are:
h i
k1 ∈ 0.25 1.75
h i (9.9)
m3 ∈ 0.20 1.80
Thus, we wish to keep z(t), the position of m2 and m3 , as small as possible in the
presence of disturbances, i.e. achieve superior disturbance-rejection.
All feedback loops utilize the two noisy measurements y(t),
(
y1 (t) = x2 (t) + θ1 (t)
(9.11)
y2 (t) = x3 (t) + θ2 (t)
that includes additive white sensor noise θ(t) = [θ1 (t) θ2 (t)]T , defined by
E{ θ(t)} = 02×2 , E{ θ(t)θ(τ )} = 10−6 δ(t − τ )I2×2
In addition to the two real uncertainties, we assume that there are unmodeled
time-delays τi in both control channels whose maximum possible values are 0.05 secs,
i.e.
τi ≤ 0.05 secs ; i = 1, 2 (9.12)
2.45s
Wun (s) = I2×2 (9.13)
s + 40
3 3
10 10
k =0.25
k1=0.25 k1=1.75 k1=0.25 1
k1=1.75 m2=1.8
m2=0.2 m2=1.8 m2=0.2
2 2 m2=1.8
10 10
k1=1.75
k1=1.75
m2=0.2
m2=0.2
1 1
10 10
k1=0.25
m2=1.8
Mag.
Mag.
0 0
10 10
−1 −1
10 10
−2 −2
10 10
−3 −3
10 10
−2 −1 0 1 −2 −1 0 1
10 10 10 10 10 10 10 10
Freq (rad/sec) Freq (rad/sec)
Figure 9.2: The singular value Bode plots of the TITO MSD system.
290
using the singular value Bode plots. Similar plots can be made for other values of the
uncertain spring and mass.
Remark: The disturbance-rejection control problem is hard even if both uncertain
parameters are known; the control problem becomes much harder in our adaptive
design, because the controls u1 (t) and u2 (t) are applied through the uncertain spring
and uncertain mass, respectively.
Remark: One of the objectives of the compensator design (GNARC and/or LNARC)
is to achieve low errors of the position x2 (t) of the mass m2 . This control problem
is also hard, since the control is not directly applied to the mass subject to the dis-
turbance (non-collocated actuator control problem). Clearly, the control problem
becomes much harder in our adaptive design, because the control u1 (t) is applied to
the mass m1 , and there is an uncertain spring in between; thus one is not sure how
much force is exerted through the uncertain spring k1 to the mass m2 .
m3 = 1.0. The weights of eqs. (9.1), (9.13), and (9.14) will not change in any of the
subsequent designs.
To carry out the mixed-µ synthesis, the parameter uncertainties of eq. (9.9) are
represented by
In order to design the “best possible” non-adaptive feedback system the following
type of performance weight upon the output vector z(t) is used:
" #
Ap 0.25 0
Wp (s) = (9.17)
s+1 0 1
which reflects our specification for good disturbance-rejection for the frequency range
ω ≤ 1 rad/sec where the disturbance d(t) has most of its power. Note that in
eq. (9.17) the first performance output, z1 (t), is selected to have four times smaller
required performance than that of the z2 (t). It means that the control accuracy of
z2 (t) is four times more important than z1 (t). We made this choice due to the fact
that the mass m3 is collocated with control u2 (t) and can be easily controlled, in
contrast with mass m2 which does not have any collocated control acting upon it.
Remark: Our design choice in eq. (9.17) to have z2 be four times more important
is subjective. Other weights could be used yielding different GNARC and RMMAC
designs.
Using the mixed-µ software [3, 47, 110, 114] the performance parameter Ap in
eq. (9.17) is maximized until the upper-bound on the mixed-µ, µub (ω), satisfies the
inequality
µub (ω) ≤ 1, ∀ω with (µub ' 0.995) (9.18)
x (t )
q (t )
1
I 2´ 2 10 -3 I 2´ 2
s +1
Plant Model Set
éd k1 0 ù W p ( s)
D un ( s ) d (t ) ê ú
ëê 0 d m3 ûú
é 0.25 ù
0 ú ~
ê s +1 z (t )
Wun ( s ) ê ú Ap
ê 0 1 ú
z (t ) ëê s + 1 ûú
+ G (s) +
~
u (t )
s + 10
I 2´ 2
s + 103
u (t ) y (t )
K (s)
Figure 9.3: MIMO weights for Robust Synthesis. Control accuracy of z2 (t) is selected
to be four times more important than z1 (t).
ξ (t) θ (t)
u (t ) y (t )
r1 (t ) u1 (t ) u (t )
r2 (t ) u2 (t )
r3 (t ) u3 (t )
r4 (t ) u4 (t )
r5 (t ) u5 ( t )
S1 ,..., S5 P1 (t ),..., P5 (t )
posterior probabilities to generate the actual adaptive control 2-vector, u(t). The
number (5) of KFs and LNARCs is the outcome of the design process described in
this section.
First, we computed the FNARC for the entire k1 − m3 plane maximizing per-
formance parameter Ap at each grid point. This is the best possible performance
achievable if both uncertain parameters were known exactly. The FNARC defines a
surface over the rectangle. The maximum Ap ’s of the FNARC (AFp ) are shown at the
corners of the rectangle in Figure 9.5.
As explained in Chapter 8, we can not generalize the SISO procedure of Chapter
5 deriving the required number of models by demanding an adaptive performance of
a chosen value of X%=22% of the FNARC. In fact, this leads to non-rectangle inter-
sections of the FNARC surfaces with the X%=22% planes that are not compatible
with the rectangle uncertainty sets required by mixed-µ synthesis.
Based on the study of the FNARC surface shapes, we decided to use the X%=22%
requirement along the top edge (m3 = 1.8) in Figure 9.6, and proceed as in the SISO
294
Figure 9.5: The best performance parameters, AFp ’s, of the FNARC found at the
corners of the 2-D uncertain rectangle.
case, i.e. the uncertainty of m3 is fixed for all local models while the uncertainty in
k1 is split for extracting the local models. For example, the Model#1 is found based
on the performance parameter of Ap = 0.22 × 312.6 = 68.77, etc. This process with
X%=22% of the FNARC resulted in the five models as in eq. (9.20).
To evaluate the RMMAC with few models, we selected the “vertical” model sub-
division as illustrated in Figure 9.6 with five models where the 2-D uncertainty space
is divided vertically.
Using the mixed-µ synthesis, five “local” robust LNARCs, K1 (s), . . . , K5 (s),
are designed for each model defined in eq. (9.20). In the mixed-µ synthesis, the
weights of eqs. (9.1), (9.13), and (9.14) are the same as in the GNARC design of
Section 9.3. However, for each LNARC design, the performance parameter, Ap , in
eq. (9.17) is again maximized until the mixed-µ bound of eq. (9.18) is achieved. Fig-
ure 9.6 also shows the “best” performance parameters (Ap ’s) used for each of the five
295
LNARC designs based on X%=22% of the FNARC. We stress that each LNARC,
K1 (s), . . . , K5 (s), is a 2 × 2 complex matrix and
converges to its nearest probabilistic neighbor. In this case, the RMMAC essentially
operates as an LTI stochastic feedback system. This allows one to calculate a key
transfer function (using sensitivity singular value plots) for disturbance-rejection in
the frequency domain
for different values of the unknown spring k1 and unknown mass m3 for the GNARC,
as in eq. (9.9), and for each LNARC design, as in eq. (9.20).
Figure 9.7 quantifies (using singular value Bode plots) the potential RMMAC
improvement in disturbance-rejection. Similar plots can be made for other values of
the uncertain spring and mass. Figure 9.7 shows that the RMMAC has the potential
to significantly improve disturbance-rejection; these predictions will be validated in
the sequel. Of course, the non-adaptive GNARC does better than the open-loop case.
We remind the reader that since the disturbance-rejection plots in Figure 9.7
involve maximum singular values, they quantify the “most pessimistic” disturbance-
rejection especially in the frequency region ω ≤ 1 rad/sec (chosen in our performance
weight, eq. (9.17)).
Figure 9.8 evaluates the potential performance improvement of using the RMMAC
in a stochastic setting, namely by comparing the RMS errors of the outputs, z1 and z2 ,
for different values of k1 and m3 . These RMS results are readily computed by solving
standard covariance algebraic Lyapunov equations for stochastic LTI systems. The
graphs of Figure 9.8 vividly suggest that RMMAC has the potential of decreasing the
RMS of output z1 by a factor of 1 to 3.5 and the RMS of output z2 by a factor of 1.65
to 4 over the GNARC system. This is a consequence of the specific weight selected,
eq. (9.17).
Finally, it is important to show the “stability-mismatch” cases shown in Figure 9.9.
The interpretation of Figure 9.9 answers the question: what happens to the closed-
loop stability if we use the LNARC#j, Kj (s), when the true spring and mass are in
Model#i? The anti-diagonal entries (rectangles) in Figure 9.9 are always robustly-
stable, by construction. Examining the top row in Figure 9.9 demonstrates that
for all k1 ∈ [1.28, 1.75] and m3 ∈ [0.20, 1.80], if we use LNARC#5 we always have
297
3 3
10 10
Open−Loop Open−Loop
2 GNARC 2 GNARC
10 10
LNARC#1 LNARC#2
σmax(Mdz(jω)
σmax(Mdz(jω)
1 1
10 10
0 0
10 10
−1 −1
10 10
−2 −2
10 10
−2 −1 0 1 2 −2 −1 0 1 2
10 10 10 10 10 10 10 10 10 10
Freq. (rad/sec) Freq. (rad/sec)
(a) k1 = 1.55 and m3 = 0.9 in M#1 (b) k1 = 1.02 and m3 = 0.9 in M#2
3 3
10 10
Open−Loop Open−Loop
2 GNARC 2 GNARC
10 10
LNARC#3 LNARC#4
σmax(Mdz(jω)
σmax(Mdz(jω)
1 1
10 10
0 0
10 10
−1 −1
10 10
−2 −2
10 10
−2 −1 0 1 2 −2 −1 0 1 2
10 10 10 10 10 10 10 10 10 10
Freq. (rad/sec) Freq. (rad/sec)
(c) k1 = 0.76 and m3 = 0.9 in M#3 (d) k1 = 0.51 and m3 = 0.9 in M#4
3
10
Open−Loop
2 GNARC
10
LNARC#5
σmax(Mdz(jω)
1
10
0
10
−1
10
−2
10
−2 −1 0 1 2
10 10 10 10 10
Freq. (rad/sec)
0.6
GNARC
0.5
RMS of Z1(t)
0.4
0.3 LNARCs
0.2
0.1
0
0
1 0.5
1
1.5
2 k1
m
3
GNARC
0.3
0.25
RMS of Z2(t)
0.2
0.15
LNARCs
0.1
0.05
2
1.5 2
1 1.5
1
0.5
0.5
m 0 0
3 k1
LNARC #1
1.8
3
1
m
0.2
LNARC #2
1.8
3 1
m
0.2
LNARC #3
1.8
3
1
m
0.2
LNARC #4
1.8
3
1
m
0.2
LNARC #5
1.8
3
1.1
m
0.2
0.25 0.5 0.75 1 1.25 1.5 1.75
k
1
M #5 M #4 M #3 M #2 M #1
Figure 9.9: Stability-mismatch analysis to see under what conditions we can have
instability if the parameter (k1 , m3 ) belongs to Model#j, but we use the LNARC for
Model#i. The ¤ (squares) denote stable pairs; the × denote unstable combinations.
instability; if we use the LNARC#3 we have instability for larger values of k1 , but
have stability for smaller values. This is due to the fact that the mixed-µ upper
bound inequality (9.18) is only a sufficient condition for both robust- and performance-
stability and, hence, each LNARC design will actually have a wider robust-stability
region. Of course, performance-robustness is only guaranteed for the models defined
in eq. (9.20).
In Section 8.3 we discussed the design of the Kalman filters (KFs) of the RMMAC
methodology in the MIMO case and provided some insight in the TITO cases. In this
section, we will follow the ideas of Section 8.3 to design the Kalman filters for the
TITO example. Note that, for any choice of a nominal (k1∗ , m∗3 ) for a KF, the BPM
defines a surface in 3-dimensional space.
300
Ξf 1 0
Ξf =
Ξf 2
0 Ξf 3
301
is as follows.
" #
h i ξ(t)
ẋ(t) = Ax(t) + Bu(t) + L Lf
ξf (t)
(9.24)
z(t) = C1 x(t)
y(t) = C x(t) + θ(t)
2
where
h iT
Lf = 03×3 I3×3 03×2 (9.25)
while the other matrices and vectors are the same as used in the normal case
(without considering fake noise) in Section 9.2.
It is again emphasized that the fake plant white noise is included for the purpose
of preventing the identification subsystem of the RMMAC, from being over-confident
in its estimates and not to switch to a different model rapidly. It is also important
to note that the Kalman filters do not know anything about the time-delays in the
system. This is an additional support to use Ξf , which increases the KF gains and
more attention is paid to the residual signals.
The following fake-white-noise intensities were found by trial and error to satisfy
the ergodicity condition, eq. 3.73.
Note the large disturbance intensity of the plant noise ξf 3 is equal to 100, and acts
upon the mass m3 . This was probably necessary because of the large m3 uncertainty
for each vertical model of Figure 9.6.
Figure 9.10 shows the BPM surfaces of Model#1 and Model#2. Note that they
intersect almost exactly at the Model#1 and #2 boundaries. For the other three
models similar surfaces are generated (not shown).
Figure 9.11 shows the final results of BPM optimization including the nominal
points of each KF and the five different models used to ensure asymptotic convergence
of the posterior probabilities, so that the RMMAC operates well and yields correct
asymptotic identification.
302
−11
BPM#1
BPM#2
BPM
−11.5
−12
0.2
0.6
1 0.895
1.4 1.28
Ω (M#1) 1.8 1.75 Ω (M#2)
1 2
m k1
3
Figure 9.10: BPM#1 and BPM#2, with the fake noise intensity of eq. (9.26).
1.8
1.4
m3
0.6
0.2
0.25 0.75 1.25 1.75
k1
Figure 9.11: Design points for each MIMO KF selected (not at the model center) so
that the BPMs for each adjacent model boundary are matched.
303
In this section, we present some stochastic simulations using the complete RMMAC
closed-loop system of Figure 9.4. Of course, testing the RMMAC requires significant
computation using multiple Monte Carlo (MC) runs under different scenarios. Due
to space limitations, only representative plots are shown; however, our conclusions
are based on many other MC runs not explicitly shown in this chapter.
Unless stated otherwise, all stochastic simulations use stochastic disturbance pro-
cesses, as defined by eq. (9.1), white measurement noises, and the true system includes
actual (but unmodeled) time-delays of 0.01 secs. The results presented in the sequel
are done for a MC simulation. The sampling period is Ts = 0.001 secs.
Figures 9.12 and 9.13 show stochastic simulations for two different cases. Figures 9.12(a)
and 9.13(a) show the five posterior probabilities, P1 (t), . . . , P5 (t), generated by the
Posterior Probability Evaluator (PPE) using different values for the uncertain mass
and spring constant. All were initialized by Pk (0) = 0.20 ; k = 1, . . . , 5.
In order to properly interpret Figure 9.12(a) and 9.13(a) the reader should com-
pare the indicated mass/spring values to the models defined in eq. (9.20). In Fig-
ures 9.12(a) we note that P4 (t) → 1, which means that Model#4 is the nearest
probabilistic neighbor. In Figure 9.13(a), P2 (t) → 1, indicating that Model#2 is the
nearest neighbor.
Figures 9.12(b)-(c) and 9.13(b)-(c) compare the two performance outputs of the
non-adaptive feedback system (GNARC) to that of the RMMAC and clearly demon-
strate the RMMAC improvement in disturbance-rejection. These results are consis-
tent with the predictions in Figure 9.8.
304
P1(t)
0
1
P2(t)
0
1
P3(t)
0
1
P4(t)
0
1
P5(t)
0
0 10 20 30
time (sec)
(a) Posterior probabilities
0.6
GNARC
0.4
RMMAC
0.2
z1(t)
0
−0.2
−0.4
0 20 40 60 80 100
time (sec)
(b) Performance output, z1
0.3
GNARC
0.2
RMMAC
0.1
z2(t)
0
−0.1
−0.2
0 20 40 60 80 100
time (sec)
(c) Performance output, z2
P (t)
1
0
1
P (t)
2
0
1
P (t)
3
0
1
P (t)
4
0
1
P (t)
5
0
0 10 20 30
time (sec)
(a) Posterior probabilities
0.4
GNARC
0.2 RMMAC
z1(t)
−0.2
−0.4
0 20 40 60 80 100
time (sec)
(b) Performance output, z1
0.4
GNARC
0.2 RMMAC
z2(t)
−0.2
−0.4
0 20 40 60 80 100
time (sec)
(c) Performance output, z2
In this set of simulations the disturbances are the sums of filtered white noises and pe-
riodic square-waves (±2) with a period of 60 secs. The sensor noises were white. The
KFs in the RMMAC were NOT aware of the square-wave disturbance; they continued
to use eq. (9.1) to model the disturbance dynamics. Thus, we have a “mild” violation
of the theoretical assumptions underlying the RMMAC. The true spring stiffness is
k1 = 0.7 and the true mass is m3 = 0.4. Figure 9.14 shows the simulation results.
Note that after about 20 secs the probabilities converge to the correct Model#3, i.e.
P3 (t) → 1 as shown in Figure 9.14(a). The RMMAC performance is very good, as
predicted.
In Figure 9.9 the mismatch-stability of the LNARC designs were shown. We next
evaluate the RMMAC response when it is forced to be unstable at time t = 30 for
∆ = 0.6 secs, i.e. 600 measurements.
307
P1(t)
0.5
0
1
P2(t)
0.5
0
1
P3(t)
0.5
0
1
P4(t)
0.5
0
1
P5(t)
0.5
0
0 10 20 30 40
time (sec)
(a) Posterior probabilities
2
d1(t)
−2
2
d2(t)
−2
−1
0 40 80 120 160 200
0.4 GNARC
0.2 RMMAC
z2(t)
0
−0.2
−0.4
0 40 80 120 160 200
time (sec)
Figure 9.14: Performance improvement by the RMMAC when the theoretical as-
sumptions are mildly violated. The value k1 = 0.7, m3 = 0.4 (in M#3) are used.
Disturbances are the sums of filtered white noises and periodic square waves (±2).
308
1
1
P1(t)
0.5 RMMAC
0.5 GNARC
0
z1(t)
1 0
P2(t)
0.5
−0.5
0
1 −1
0 50 100 150 200
P3(t)
0.5 GNARC
0.5
z2(t)
0
1 0
P5(t)
0.5
0 −0.5
0 50 100 150 200 0 50 100 150 200
time (sec) time (sec)
Figure 9.15: Simulation results of the RMMAC performance when the actual distur-
bances d(t) are sinusoidal disturbances, with period of 60 secs, superimposed with
the original stochastic disturbances. The values k1 = 1.2 and m3 = 1.7 in Model#2
(and also close to Model#1) are used. Only one MC run is shown.
Figure 9.16 illustrates a typical result selected from several different MC sim-
ulations. In Figure 9.16 the true value of k1 = 1.5 and m3 = 1.0 (in Model#1)
⇒ k1 ∈ [1.28, 1.75]; its nearest probabilistic neighbor is Model#1. From Fig-
ure 9.9 we know that if we use LNARC#5 we have an unstable closed-loop sys-
tem. To force this instability the initial values of the probabilities vector are equal
(Pj (0) = 0.2 ; j = 1, . . . , 5), but at t∗ =30 secs
so that at t = 30 secs the RMMAC system is forced to be unstable for 0.6 secs.
However, as illustrated in Figure 9.16, the RMMAC rapidly recovers to a stable
configuration. Figure 9.16(a) shows that the “correct probability” P1 (t) → 1 quickly,
starting from its forced value P1 (30) = 0.01; the other four probabilities converge to
zero quickly as well. Figure 9.16(b) shows the output response in which, after a small
period of slight oscillation, after t = 30 secs, the RMMAC recovers, converges to
309
the right model, i.e. P1 (t) → 1, and returns to its predictable superior disturbance-
rejection.
P1(t) 1
0.5
0
1
P2(t)
0.5
0
1
P3(t)
0.5
0
1
P4(t)
0.5
0
1
P5(t)
0.5
0
0 20 40 60
time (sec)
(a) Posterior probabilities
0.8
0.4
z1(t)
−0.4
RMMAC
GNARC
−0.8
0 40 80 120 160 200
time (sec)
0.4
0.2
z2(t)
−0.2
RMMAC
LNARC
−0.4
0 40 80 120 160 200
time (sec)
Figure 9.16: RMMAC recovering from forced instability at t = 30 secs. The values
k1 = 1.5, m3 = 1.0 (in M#1) is used.
311
1.8
1.4 Start
m3
0.6
0.2
M#5 M#4 M#3 M#2 M#1
0.25 0.5 0.75 1 1.25 1.5 1.75
k
1
1 GNARC
RMMAC
0.5
0
−0.5
−1
0 500 1000 1500 2000 2500 3000 3500 4000
0.5
GNARC
RMMAC
−0.5
0 500 1000 1500 2000 2500 3000 3500 4000
time (sec)
Here, the plant disturbances used have the square-wave intensities with two intensity
values, Ξ1 = 1 and Ξ2 = 100 and with a period of 50 secs as shown in Figure 9.18(a).
The spring constant k1 = 0.3 and the mass m3 = 0.3 are used (in Model#5).
The posterior probabilities of the RMMAC are shown in Figure 9.18(b). In this
case, as to be expected, the RMMAC behaves “poorly” and probability oscillations
are observed during the times that the disturbance are very large. The KFs were
designed using Ξ = 1. We shall discuss next how to overcome these shortcomings,
due to unknown plant disturbances, using the RMMAC/XI architecture discussed in
Section 5.7. The performance outputs are shown in the sequel.
1
P1(t)
0.5
0
1
P2(t)
0.5
0
1
P3(t)
0.5
100 0
1
P4(t)
0.5
Ξ(t)
1
P5(t)
0.5
0
1 0 50 100 150 200
0 50 100 150 200
time (sec) time (sec)
(a) Square-wave intensity of plant disturbances (b) Posterior probabilities transients. Pi (0) = 0.2
; i = 1, . . . , 5.
Figure 9.18: The RMMAC behavior for varying plant noises intensity. The spring
constant k1 = 0.3 and the mass m3 = 0.3 are in Model#5. The KFs were designed
using Ξ = 1.
Then we design two sets of linear KFs, one set indexed by k = 1, . . . , 5 using the
313
noise intensity Ξ1 = 1 and the second set indexed by k = 6, . . . , 10 using the noise
intensity Ξ2 = 100. We use the same nominal points for designing each KF in the
second set as in Table 9.1.
The simulation results are shown in Figure 9.19. The intensity of each plant dis-
turbances is the same shown in Figure 9.18(a). Recall that only correct probabilities
for k1 = 0.3 and m3 = 0.3 in M#5, i.e. P1 (t) and P6 (t), are selected that are re-
spectively associated with the KF#1 (designed with Ξ = 1) and the KF#5 (designed
with Ξ = 100). Other posterior probabilities are almost zero. Figures 9.19(a) and
(b) show that the RMMAC/XI system corrects the performance degradation of the
standard RMMAC.
9.8 Conclusions
Based on the simulations presented in this chapter (and many others not shown due
to space limitations) we have demonstrated that the TITO RMMAC design works
very well.
As to be expected, the MIMO off-line design is more complex and demanding
than the SISO one. This increase in difficulty arises in the systematic definition
of the minimum number of models using the FNARC% method, as well as in the
definition of the nominal parameter values to be used for designing the KFs.
In view of the relatively large parametric uncertainty in each model (the entire m3
axis), we found it necessary to “robustify” the KFs using the time-honored engineering
technique that includes ”fake white process noise” for each KF. This has the effect that
the KF gains are increased and more attention is paid to the data. The benefits are
better convergence of the posterior probabilities to the nearest probabilistic neighbor
and, hence, correct asymptotic identification. Without the use of fake white process
noise we could not arrive at the result shown in Figure 9.11 regarding the location of
the KF nominal design parameters.
When the RMMAC did not violate any of the theoretical assumptions, it worked
extremely well and validated the performance predictions; see Figures 9.12 and 9.13.
314
12 5
RMMAC/XI RMMAC/XI
10 RMMAC RMMAC
4
8
3
6
2
4
z1(t)
z2(t)
2 1
0
0
−2
−1
−4
−2
−6
−8 −3
0 50 100 150 200 0 50 100 150 200
time (sec) time (sec)
(a) The performance output z1 (t) (b) The performance output z2 (t)
1 1
P1(t)
P6(t)
0.5 0.5
0 0
1 1
P2(t)
P7(t)
0.5 0.5
0 0
1 1
P3(t)
P8(t)
0.5 0.5
0 0
1 1
P4(t)
P9(t)
0.5 0.5
0 0
1 1
P10(t)
P5(t)
0.5 0.5
0 0
0 50 100 150 200 0 50 100 150 200
time (sec) time (sec)
Figure 9.19: Comparison of the standard RMMAC and the RMMAC performance
for varying plant noises intensity as in Figure 9.18(a). The spring constant k1 = 0.3
and the mass m3 = 0.3 are in Model#5 with Ξ = 1 and in Model#10 with Ξ = 100.
Note that the correct probabilities, P5 (t) and P10 (t), (almost immediately) go to unity
when Ξ changes.
315
When the RMMAC violated the theoretical assumptions in a mild way, see Fig-
ure 9.14, it still delivered very good performance. As shown in Figure 9.16, it even
recovered rapidly from an externally imposed brief instability. It also worked quite
well for the slowly time-varying parameter case, see Figure 9.17. This is very im-
portant, because one of the major reasons for using adaptive control is to be able
to deliver superior performance in the case of slowly changing parameters. These
conclusions are reinforced by many other simulations not shown in this thesis.
We also “pushed” the RMMAC designs by applying disturbances of unknown
intensity (representing severe violations of the theoretical assumptions) so as to isolate
and expose the RMMAC potential weaknesses. The large differences between the
design and the actual plant disturbance intensities may lead to rapid switching of
the posterior probabilities, poor performance, and even instability. Thankfully, the
RMMAC/XI architecture seems to solve such these problems, at the price of using
two sets of the Kalman filters in the RMMAC architecture.
In our computer, the simulation time of the RMMAC with five LNARCs/KFs –
see Figure 9.4 – for a single MC run, with a sampling time of Ts = 0.001 secs, was
approximately 1.5 times faster than real-time.
316
Chapter 10
In this chapter, we summarize the most relevant results of the results of this thesis in
Section 10.1. We then suggest, in Section 10.2, an “easy” extension to the RMMAC
methodology; this is the direct consequence of some of our theoretical and simulation
investigations. Finally, in Section 10.3, we list numerous specific suggestions for future
theoretical and applied research motivated by our research experience in this thesis.
317
318
for feedback design, namely unmodeled dynamics, one or more uncertain real pa-
rameters, unmeasurable plant disturbances and unmeasurable sensor noise, together
with suitable modeling of the exogenous signals and explicit quantitative performance
specifications.
(4). The RMMAC methodology makes transparent the tradeoffs between the
complexity of the multiple-model adaptive control system and the posed quantitative
performance requirements. In general, to get better and better performance one must
use RMMACs with increased complexity; on the other hand, the point of diminishing
returns in this tradeoff also becomes transparent through the use of the FNARC.
(5). In the case of a single scalar uncertain parameter, the RMMAC methodology
provides a very detailed step-by-step design process for the design of the multiple-
model adaptive identification and control subsystems. For two or more uncertain
parameters the step-by-step process still holds but the detailed designs become much
more complex, as to be expected. However, RMMAC can readily accommodate
several controls and measurements, i.e. it is a truly multivariable adaptive design
unlike 99% of the results presented in the literature to-date.
(6). We have evaluated, via thousands of simulations and test cases, the RMMAC
performance and have compared it to that of the best nonadaptive design (GNARC).
The simulation results presented in this thesis show excellent RMMAC performance
when the theoretical assumptions are not violated and very good performance when
the theoretical assumptions are “mildly” violated (including the case of time-varying
parameters). We also showed that the RMMAC can behave poorly, and even break into
instability, when the assumptions were “severely” violated1 ; in these cases we have
identified the mechanisms that lead into this undesirable RMMAC behavior and we
suggested some specific directions to overcome these shortcomings. The RMMAC/XI
architecture discussed in Section 5.7 and evaluated in Section 6.3.7, for the LFD case,
Section 6.4.5, for the HFD case, and in Section 9.7.5 for the TITO case represent
such improvements.
Has the RMMAC methodology led us to the “pot of gold at the end of the adaptive
rainbow”? The answer is NO; perhaps, we are a bit closer. Much more research is
1
Theories have limitations; stupidity does not!
319
needed to approach the elusive point. A critical missing element is the lack of global
stability-theory results that can address the highly nonlinear, time-varying, stochastic
characteristics of the dynamic feedback system associated with any RMMAC design.
ξ (t) θ (t)
u (t ) y (t )
r1 (t ) u1 (t )
S1 (t ) K1 ( s )
r2 (t ) u2 (t ) u (t )
S 2 (t ) K 2 ( s)
rN (t )
u N (t )
S N (t ) K N (s)
P1 (t )
P2 (t )
PN (t )
as random vectors and are appended to the plant state vector. The initial covariance
of the uncertain parameters (its 3-sigma level) will be estimated from the size of the
associated sub-model rectangle. In this case, we deal with a nonlinear estimation
problem and the use of the EKFs is perfectly natural. Indeed, the multiplicity of
models is consistent with the Sum of Gaussians method, Refs. [5, 16, 53, 54], that has
been widely used in high-performance nonlinear estimation applications.
Clearly, the on-line calculations required for the bank of the EKFs will be more
extensive, since the EKF filter gain-matrices must be calculated on-line. Also, we
must calculate the residual covariance matrices, Sk (t), in real-time. The net outcome
of this would be the generation of the EKF residuals, rk (t),which are fed together with
the residual covariance matrices,Sk (t), to the PPE which now utilizes the time-varying
MMAE version calculations (see Section 3.3.3) to obtain the posterior probabilities,
321
repeated below.
βk (t + 1)e−(1/2)wk (t+1)
Pk (t + 1) = · Pk (t) (10.1)
P
N
Pj (t)βj (t + 1)e−(1/2)wj (t+1)
j=1
where
1
βi (t + 1) = p (10.2)
(2π)m/2 det Si (t + 1)
wi (t + 1) = ri0 (t + 1)Si−1 (t + 1)ri (t + 1) (10.3)
be examined including multivariable versions in which, for example, the TITO MSD
system of Chapter 9 can serve as the initial testbed.
Most certainly, comparisons using systems that have at least one uncertain pole
and uncertain zero in the right half s-plane must be carried out. We know that for
such unstable and nonminimum-phase plants there are severe limitations, Refs. [108,
125, 127], in designing robustly stable systems with superior performance even in the
non-adaptive case. Some of the issues in dealing with a nonminimum-phase zero
plant were discussed in Chapter 7. Adding an unstable uncertain pole will greatly
complicate matters. We believe that a great deal of care must be exercised in defining
any MMAC and, of course, fair RMMAC and SMMAC comparisons would be most
valuable.
It is not at all clear that we can compensate for these unmodeled dynamics by
using “fake plant white noise” in the KF designs; at best, this will require trial-and-
error. Rather we need a more complete theory that extends the results of [44] to
account for unmodeled dynamics. One of the potential difficulties will be the fact
that the MMAE convergence results are derived purely in the time-domain, while
unmodeled dynamics are bounded in the frequency-domain. On the other hand, the
presence of LTI unmodeled dynamics should not destroy the asymptotic ergodicity
and stationarity assumptions in Ref. [44].
324
Throughout this thesis we have relied heavily upon the MMAE asymptotic conver-
gence results using the “Baram Proximity Measure (BPM)”, [44, 45, 52, 99], as dis-
cussed in Chapter 3. As we have remarked, these asymptotic convergence results can
only be proven under certain ergodicity and stationarity assumptions. We remind
the reader that these results are only available in the open-loop case. For the closed-
loop RMMAC system, if the stochastic disturbances are stationary and if indeed,
after a transient interval, the posterior probabilities converge to a model, then the
adaptive control signal approaches a stationary process and we have postulated that
convergence will occur and be maintained.
However, if we mix stochastic and deterministic plant disturbances (as we did in
Sections 6.3.5.2, 6.4.4.2, 7.5.2, and 9.7.2), where we added square-waves or sinusoids
to the filtered white noise plant disturbance or if the uncertain parameter(s) were
time-varying (as we did in Sections 6.3.5.2, 6.3.5.3, 7.5.2.4, and 9.7.4), then we are
certainly violating the ergodicity and stationarity conditions. Also, if during the
initial transient period there is frequent, but temporary, “switching” of the posterior
probabilities, then we can no longer make a claim that the adaptive control signal is
stationary. Nonetheless, our simulation results showed that the RMMAC still works
quite well. There must be an underlying theoretical reason for these positive and
encouraging empirical observations. We, therefore, believe that a fresh theoretical
approach that extends and generalizes the original results of Y. Baram [44] is a most
worthwhile avenue for further theoretical investigations.
In this thesis we utilized the BPM (see Section 3.5) to quantify the distance between
different models. Indeed, the use of the BPM was critical in designing the bank of KFs
in the RMMAC so as to match the BPMs at the model boundaries; see Sections 5.3
325
and 8.3. On the other hand, other researchers have utilized the Vinnicombe metric,
[39, 131, 132], as a (deterministic) measure of distance among models utilized in the
SMMAC architecture.
In certain of our numerical studies, not explicitly presented in this thesis, we
observed for some examples a great deal of similarity between the BPM and the
Vinnicombe metric. We do not know if this was a coincidence or whether or not
there is a deeper relation between them. This is another topic for future theoretical
and simulation research.
The results of Chapters 8 and 9 clearly indicate that it becomes extremely complicated
to design the RMMAC KFs and LNARCs when there are three or more uncertain real
parameters. We suspect that these difficulties will also be present for any extension
of the SMMAC architectures, which currently are documented only for a single scalar
parameter.
In such multiple uncertain parameter cases, the systematic % FNARC method
for determining the number of models is very complex or breaks down. Also, the
determination of the nominal design points for the KFs, so as to match the BPM at
the model boundaries, is also extremely burdensome. Even the performance of the
mixed-µ software becomes increasingly complex, because the underlying optimization
problems are NP-complete and nonconvex (with an increasing multiplicity of possible
local minima as the number of uncertain parameters increases).
In the mixed-µ methodology the design of the robust compensator critically de-
pends on the performance weights. It is reasonable to expect, from an engineering
perspective, that for any given performance weights perhaps only one or two of the
uncertain parameters are “dominant”, or most critical, especially after the order of
the robust compensator has been reduced. However, we are not aware of a method-
ology – with the exception of the recent work of Ref. [121] – that can be used to
isolate these one or two “dominant” parameters and compensates, perhaps, for the
remaining ones by structured complex-valued uncertainty. Indeed, if we change the
326
All adaptive methodologies to-date, including RMMAC, have been developed for
constant uncertain parameters. Some have been subsequently tested for time-varying
parameters, as we did in Sections 6.3.5.2, 6.3.5.3, 7.5.2.4, and 9.7.4; see also [32].
Recently, a significant amount of research has been devoted to the problem of
Linear Parameter Varying (LPV) systems; see, e.g. [135–138]. Some research on LPV
systems was initially motivated [91] by the lack of theory associated with engineer-
ing “gain-scheduled” designs [92, 93, 139–142]. Although at present a complete LPV
design methodology (and software) is not available it may be worthwhile to attempt
to integrate the RMMAC concepts and the LPV results in future theoretical studies
since both attempt to address similar problems.
When we violate the theoretical assumptions in a severe way, then the RMMAC
can behave poorly and even become unstable. In Chapters 6 and 9 we discussed the
“abnormal” behavior of the residual signals in these situations. Some problems can be
fixed using “fake plant noise”, and some by the RMMAC/XI architecture. However,
it is important to introduce some additional “intelligence” in any adaptive design
that first senses that the adaptive system does not operate normally and second to
“turn-it-off”, either temporarily or for the remainder of the operation.
In the RMMAC design we can carry out a variety of the real-time tests involving
the residual signals, by testing for abnormal magnitudes, lack of stationarity and
ergodicity etc. We can then use these “abnormality tests” to cause the adaptive
system be disconnected from the feedback loop and the GNARC be connected in the
feedback loop. By design, the GNARC will provide robust-stability, albeit with poorer
performance. While the GNARC is operating, the MMAE residuals and posterior
probabilities can continue to be monitored; if they return to a more-or-less “normal”
environment then the GNARC could be disconnected and the RMMAC reconnected.
Needless to say, there are a host of unresolved signal processing issues that must be
researched and resolved before such “fail-safe” logic can be streamlined, implemented
and tested.
328
Another unresolved issue is when the plant is both unstable and highly uncertain
so that a single GNARC does not exist. This would imply that we must relax the
performance specifications.
[2] ——, “Computing bounds for the mixed-µ problem,” Int. J. of Robust and
Nonlinear Control, vol. 5, pp. 573–590, 1995.
[3] G. Balas et al., µ-Analysis and Synthesis Toolbox of Matlab, User’s Guide,
4th ed. The MathWorks Inc., June 2001.
[5] P. Maybeck, Stochastic Models, Estimation and Control. Academic Press, 1979
and 1982, vol. I and II.
[6] J. Krause et al., “Robustness studies in adaptive control,” in Proc. of the IEEE
Conf. on Decision and Control, Dec. 1983, pp. 977–981.
[8] S. Sastry and M. Bodson, Adaptive Control: Stability, Convergence and Ro-
bustness. NJ, USA: Englewood Cliffs, Prentice-Hall, 1989.
[9] P. Ioannou and J. Sun, Robust Adaptive Control. NJ, USA: Prentice-Hall,
1996.
[10] I. Landau, Adaptive Control: The Model Reference Approach. NY, USA: Mar-
cel Dekker, 1979.
329
330
[11] K. Narendra and A. Annaswamy, Stable Adaptive Systems. NJ, USA: Prentice-
Hall, 1988.
[12] K. Åström and B. Wittenmark, Adaptive Control, 2nd ed. MA, USA: Addison-
Wesley, 1995.
[13] A. Packard and J. Doyle, “The complex structured singular value,” Automatica,
vol. 29, no. 1, pp. 71–109, 1993.
[14] G. Balas and A. Packard, The Structured Singular Value (µ) Framework, ser.
The Control Handbook, (W.S. Levine, ed.). CRC Press Inc, 1996, pp. 671–687.
[16] B. Anderson and J. Moore, Optimal Filtering. NJ, USA: Englewood Cliffs,
Prentice-Hall, 1979.
[17] K. Schott and B. Bequette, Multiple Model Adaptive Control, ser. Multiple Model
Approaches to Modeling and Control, (R. Murray-Smith and T.A. Johnsen eds.).
London: Taylor & Francis, 1997, ch. 11, pp. 269–292.
[18] M. Athans et al., “The stochastic control of the F-8C aircraft using the multiple
model adaptive control (MMAC) method – part I: Equilibrium flight,” IEEE
Trans. on Automatic Control, vol. 22, no. 5, pp. 768–780, Oct. 1977.
[19] P. Maybeck and R. Stevens, “Reconfigurable flight control via multiple model
adaptive control methods,” IEEE Trans. on Aerospace and Electronic System,
vol. 27, no. 3, pp. 470–480, 1991.
[21] P. Maybeck and J. Griffin, “MMAE/ MMAC control for bending with multi-
ple uncertain parameters,” IEEE Trans. on Aerospace and Electronic System,
vol. 33, no. 3, pp. 903–912, 1997.
[23] ——, “A new robust adaptive control method using multiple-models,” in Proc.
12th IEEE Mediterranean Conference on Control and Automation (MED’04),
Kusadasi, Turkey, June 2004.
[24] ——, “RMMAC: A novel robust adaptive control scheme – Part I: Architec-
ture,” in Proc. of the IEEE Conf. on Decision and Control, Paradise Island,
Bahamas, Dec. 2004, pp. 1134–1139.
[26] H. Tjabyadi et al., “Vibration control of a cantilever beam using multiple model
adaptive control,” in Proc. of the American Control Conf., vol. 3, June 30-July
2 2004, pp. 2907–2908.
[28] R. Middleton et al., “Design issues in adaptive control,” IEEE Trans. on Au-
tomatic Control, vol. 33, no. 1, pp. 50–58, Jan. 1988.
[30] T. Johansen and R. Murray-Smith, The operating regime approach, ser. Mul-
tiple Model Approaches to Modeling and Control, (R. Murray-Smith and
T.A. Johnsen, eds.). UK: Taylor & Francis, 1997, ch. 1, pp. 3–72.
[31] R. Kosut and B. Anderson, “Adaptive control via finite modeling and robust
control,” IFAC Workshop on Robust Adaptive Control, pp. 91–95, Aug. 1988.
[34] ——, “Adaptive control using multiple models,” IEEE Trans. on Automatic
Control, vol. 42, no. 2, pp. 171–187, Feb. 1997.
[36] J. Fitch and P. Maybeck, “Multiple model adaptive control of a large flexible
space structure with purposeful dither for enhanced identifiability,” in Proc. of
the IEEE Conf. on Decision and Control, vol. 3, 1994, pp. 2904–2909.
[37] B. Anderson et al., “Multiple model adaptive control. part 1: Finite controller
coverings,” Int. J. of Robust and Nonlinear Control, vol. 10, no. 11-12, pp.
909–929, 2000.
[38] J. Hespanha et al., “Multiple model adaptive control – part 2: switching,” Int.
J. of Robust and Nonlinear Control, vol. 11, no. 5, pp. 479–496, Apr. 2001.
[39] B. Anderson et al., “Multiple model adaptive control with safe switching,” Int.
J. of Adaptive Control and Signal Processing, vol. 15, no. 5, pp. 445–470, Aug.
2001.
333
[40] M. Safonov and T.-C. Tsao, The Unfalsified Control Concept: A Direct Path
From Experiment To Controller, ser. Feedback Control, Nonlinear Systems, and
Complexity, (B.A. Francis and A.R. Tannenbaum, eds.). Springer-Verlag, 1995,
pp. 196–214.
[41] ——, “The unfalsified control concept and learning,” IEEE Trans. on Automatic
Control, vol. 42, no. 6, pp. 843–847, June 1997.
[45] Y. Baram and N. Sandell, “An information theoretic approach to dynamical sys-
tems modeling and identification,” IEEE Trans. on Automatic Control, vol. 23,
no. 1, pp. 61–66, Feb. 1978.
[46] P. Young et al., “µ analysis with real parametric uncertainty,” in Proc. of the
IEEE Conf. on Decision and Control, 1991, pp. 1251–1256.
[47] P. Young, “Controller design with mixed uncertainties,” in Proc. of the Amer-
ican Control Conf., Baltimore, MD, June 1994, pp. 2333–2337.
[50] C. Chang and M. Athans, “State estimation for discrete systems with switching
parameters,” IEEE Trans. on Aerospace and Electronic System, vol. 14, no. 3,
pp. 418–425, May 1978.
[52] Y. Baram and N. Sandell, “Consistent estimation on finite parameter sets with
application to linear systems identification,” IEEE Trans. on Automatic Con-
trol, vol. 23, no. 3, pp. 451–454, June 1978.
[54] D. Alspach and H. Sorenson, “Nonlinear baysian estimation using gaussian sum
approximations,” IEEE Trans. on Automatic Control, vol. AC-17, pp. 439–448,
Aug. 1972.
[56] A. Gelb, Applied Optimal Estimation. MA, USA: MIT Press, 1974.
[58] H. Sorenson (ed.), Kalman Filtering: Theory and Applications. IEEE Press,
1985.
[60] H. Kwakernaak and R. Sivan, Linear Optimal Control Systems. NY, USA:
John Wiley, 1972.
335
[62] B. Anderson and J. Moore, Optimal Control - Linear Quadratic Methods. NJ,
USA: Englewood Cliffs, Prentice-Hall, 1990.
[63] L. Lublin et al., H2 (LQG) and H∞ control, ser. The Control Handbook,
(W.S. Levine ed.). CRC Press Inc, 1996, pp. 651–661.
[64] E. Jonckheere and L. Silverman, “Spectral theory of the linear quadratic optimal
control problem: discrete-time single-input case,” IEEE Trans. on Circuits.
Syst., vol. CAS-25, no. 9, pp. 810–825, 1978.
[65] C. Greene, “An analysis of the multiple model adaptive control algorithm,”
Ph.D. dissertation, MIT, Cambridge, MA, USA, 1978.
[66] C. Greene and A. Willsky, “An analysis of the multiple model adaptive control
algorithm,” in Proc. of the IEEE Conf. on Decision and Control, 1980, pp.
1142–1145.
[67] H. Shomber, “An extended analysis of the multiple model adaptive control
algorithm,” Ph.D. dissertation, MIT, Cambridge, MA, USA, 1980.
[68] Y. Baram and D. Eidelman, “An information approach to fixed gain design,”
IEEE Trans. on Aerospace and Electronic System, vol. 21, no. 1, pp. 47–55,
Jan. 1985.
[69] S. Sheldon and P. Maybeck, “An optimizing design strategy for multiple model
adaptive estimation and control,” IEEE Trans. on Automatic Control, vol. 38,
no. 4, pp. 651–654, Apr. 1993.
[71] J. Doyle, “Guaranteed margins for LQG regulators,” IEEE Trans. on Auto-
matic Control, vol. 23, no. 4, pp. 756–757, Aug. 1978.
[75] K. Narendra et al., “Adaptation and learning using multiple models, switching,
and tuning,” IEEE Control Systems Magazine, vol. 15, no. 3, pp. 37–51, June
1995.
[77] J. Zhang and D. Seborg, “Adaptive control based on switching among a set of
candidate controllers,” AICHE Annual Conference, Nov. 2001.
[79] ——, A bound for the disturbance - to - tracking - error gain of a supervised
set point control system, ser. Perspectives in Control Theory and Applications,
(D. Normand-Cyrot, ed.). Berlin: Springer-Verlag, 1998, pp. 23–41.
[82] T. Fujinaka and S. Omatu, “A switching scheme for adaptive control using
multiple models,” in IEEE Int. Conf. on Systems, Man and Cybernetics, vol. 5,
1999, pp. 80–85.
[84] F. Pait and F. Kassab, “On a class of switched, robustly stable, adaptive sys-
tems,” Int. J. of Adaptive Control and Signal Processing, vol. 15, pp. 213–238,
2002.
[85] Y. Zhang and L. Guo, “Stochastic adaptive switching control based on multiple
models,” 15th IFAC World Congress, pp. 1542–1547, July 2002.
[88] ——, “Issues, progress, and new results in robust adaptive control,” to be
appear in the International Journal of Adaptive Control and Signal Processing,
2006.
[89] M. Athans, S. Fekri and A. Pascoal, “Issues on robust adaptive feedback con-
trol,” in Preprints 16th IFAC World Congress, Invited plenary paper, Prague,
Czech Republic, July 2005, pp. 9–39.
[91] J. Shamma, “Analysis and design of gain scheduled control systems,” Ph.D.
dissertation, Dep. Mech. Eng., M.I.T., Cambridge, 1988.
[92] J. Shamma and M. Athans, “Analysis of gain scheduled control for nonlinear
plants,” IEEE Trans. on Automatic Control, vol. 35, no. 8, pp. 898–907, Aug.
1990.
[93] W. Rugh, “Analytical framework for gain scheduling,” IEEE Control Systems
Magazine, vol. 11, no. 1, pp. 79–84, Jan. 1991.
[94] R. Brown and P. Hwang, Introduction to random signals and applied Kalman
filtering: with Matlab exercises and solutions, 3rd ed. New York: John Wiley
& Sons, 1997.
[95] M. Grewal et al., Global Positioning Systems, Inertial Navigation, and Integra-
tion. John Wiley & Sons, Inc., 2001.
[98] Y. Baram and N. Sandell, “Consistent estimation on finite parameter sets with
application to dynamic system identification,” Electronic Systems Laboratory,
MIT, Cambridge, MA, USA, Tech. Rep. ESL-P-699, July 1976.
[99] K. Yared, “Maximum likelihood identification of state space models for linear
dynamic systems,” Ph.D. dissertation, MIT, Cambridge, MA, USA, 1979.
[100] S. Kullback, Information Theory and Statistics. New York: John Wiley and
Sons, Inc., 1959.
[101] J. Doob, Stochastic Processes. New York: John Wiley & Sons, Inc., 1953.
[102] P. Halmos, Lectures on Ergodic Theory. New York: Chelsea Publ. Co., 1956.
339
[104] U. Grenander, Stochastic Processes and Statistical Inference. Arkiv für Matem-
atik, 1950, vol. 1.
[107] K. Zhou et al., Robust and Optimal Control. NJ, USA: Englewood Cliffs,
Prentice-Hall, 1996.
[109] J. Doyle, “Synthesis of robust controllers and filters,” in Proc. of the IEEE
Conf. on Decision and Control, 1983, pp. 109–124.
[110] K. Zhou and J. Doyle, Essentials of Robust Control. NJ, USA: Englewood
Cliffs, Prentice-Hall, 1998.
[113] G. Stein and J. Doyle, “Beyond singular values and loopshapes,” AIAA Journal
of Guidance and Control, vol. 14, no. 1, pp. 5–16, Jan. 1991.
[115] M. Fan et al., “Robustness in the presence of mixed parametric uncertainty and
unmodeled dynamics,” IEEE Trans. on Automatic Control, vol. 36, no. 1, pp.
25–38, Jan. 1991.
340
[117] P. Young and J. Doyle, “Computation of µ with real and complex uncertainties,”
in Proc. of the IEEE Conf. on Decision and Control, 1990, pp. 1230–1235.
[119] O. Toker and H. Özbay, “On the NP-hardness of the purely complex µ com-
putation, analysis/synthesis, and some related problems in multi-dimensional
systems,” in Proc. of the American Control Conf., 1995, pp. 447–451.
[120] C. Lawrence et al., “A fast algorithm for the computation of an upper bound
on the µ-norm,” Automatica, vol. 36, pp. 449–456, 2000.
[122] A. Megretski, “Lecture notes on model order reduction,” Dept. of Elec. Eng.
and Computer Science, MIT, April 2001.
[124] P. M. Young, “Structured singular value approach for systems with parametric
uncertainty,” Int. J. of Robust and Nonlinear Control, vol. 11, pp. 653–680,
2001.
[127] J. Freudenberg and D. Looze, “Right-half plane zeros and poles and design
tradeoffs in feedback systems,” IEEE Trans. on Automatic Control, vol. 30, pp.
555–565, 1985.
[128] ——, Frequency domain properties of scalar and multivariable feedback systems.
Springer Verlag, 1988.
[132] J. Paxman and G. Vinnicombe, “Optimal transfer schemes for switching con-
trollers,” in Proc. of the IEEE Conf. on Decision and Control, Dec. 2000, pp.
1093–1098.
[133] T. Georgiou, “On the computation of the gap metric,” in Proc. of the IEEE
Conf. on Decision and Control, Austin, TX, Dec. 1988, pp. 1360–1361.
[134] M. Vidyasagar, “The graph metric for unstable plants and robustness estimates
for feedback stability,” IEEE Trans. on Automatic Control, vol. 29, no. 5, pp.
403–418, May 1984.
[135] J. Bokor and G. Balas, “Linear parameter varying systems: A geometric the-
ory and applications,” in Preprints 16th IFAC World Congress, Prague, Czech
Republic, July 2005, pp. 69–79.
342
[136] X.-D. Sun and I. Postlethwaite, “Affine LPV modeling and its use in gain-
scheduled helicopter control,” UKACC Int. Conf. on Control (Conf. Publ. No.
455), vol. 2, pp. 1504–1509, Sep. 1998.
[137] G.I. Bara et al., “State estimation for affine LPV systems,” Proc. of the IEEE
Conf. on Decision and Control, vol. 5, pp. 4565–4570, 2000.
[138] I. Szszi et al., “LPV detection filter design for a Boeing 747-100/200 aircraft,”
AIAA Journal of Guidance, Dynamics and Control, vol. 18, no. 3, pp. 461–470,
2005.
[139] R. Nichols et al., “Gain scheduling for H∞ controllers: A flight control exam-
ple,” IEEE Trans. on Control Systems Technology, vol. 1, no. 2, pp. 69–79, June
1993.
[141] F. Blanchini, “The gain scheduling and the robust state feedback stabilization
problems,” IEEE Trans. on Automatic Control, vol. 45, no. 11, pp. 2061–2070,
Nov. 2000.