Professional Documents
Culture Documents
1
ISSN 1974-6059 January 2017
Abstract – The competition of connected sources to access to the resources within a network
leads to a congestion that causes a global delay in the network. For this reason, it's challenging to
design an optimized controller that stabilizes the system and reduces the present delay. This paper
proposed a developed PID (Proportional Integral Derivative) controller based on an extension of
Hermite-Biehler theorem applicable to quasipolynomials, that is to say to systems with delay.
Then the stability region of the PID controller parameters was obtained. An improved Multi-
objective Genetic Algorithm (GA) is employed to seek the optimal PID controller gains such that
performance indices of integrated-absolute error (IAE), integrated-squared error (ISE),
integrated-time-absolute error (ITAE) and integrated-time-squared error (ITSE) are minimized,
and thereby a stability of TCP (Transmission Control Protocol)network is guaranteed. The
performance of the proposed control scheme is evaluated via a series of numerical simulations
that show its efficiency. Copyright © 2017 Praise Worthy Prize S.r.l. - All rights reserved.
Copyright © 2017 Praise Worthy Prize S.r.l. - All rights reserved DOI: 10.15866/ireaco.v10i1.11143
33
Samira Chebli, Ahmed Elakkary, Nacer Sefiani
these regulators have failed to find the whole set of discontinuities of the discrete model by making
stability values. continuous evolutions like the flow of a fluid. Inspired by
This research paper comes to fill this gap and provide the dynamics observed, the principle is to deduce
a satisfactory solution to determine the stability region intuitive equations. Once the derived equations are
delimited by the parameters of the PID controller. Based written, their dynamics are often studied using the
on the work of Silva et al, we will use the Hermite- control theory [15] – [20].
Biehler's theorem approach extended to
quasipolynomials, as a criterion of stability.
The principle of this approach lies in the fact that the Bottleneck
study of the stability of these systems controlled in TCP Link TCP
closed loop generally returns to the study and the Senders Receivers
analysis of the roots of the associated characteristic
equation.
It is no longer a matter of solving a polynomial
equation in Laplace variable but of solving a Fig. 1. Topology studied
polynomial equation having as variables and , where
Fig.1.
The non-linear Topology
model studied
used for the study is given by
is the delay. This equation, which is known as quasi-
polynomial, has generally an infinity of roots because of the following two coupled non-linear differential
the presence of the term in its expression, which equations:
implies that the stability of the system depends on the
location of these roots in the complex plan. In the case of
delayed systems, the number of roots with a positive real
part is finite. (1)
This theorem contributed to the determination of a
closed loop stability domain of a first order delay system
controlled by P, PI and PID regulators [11, 12]. Indeed,
the strong point of this technique that it remains where .
applicable to unstable systems in open loop whereas Since linear systems are easier to understand and are
several other methods of synthesis of controllers based necessary to perform a stability analysis and simulation,
on a principle of the predictor (Smith predictor, we linearize the system (1) around an equilibrium point
placement of spectrum ...) are only applicable in the case defined by:
of linear systems with stable open-loop delays.
Once the set of stabilizing controller parameters is
determined, the selection of the optimal ones is required. (2)
For this purpose, we use Genetic Algorithms (GA) as an
optimization tool. This optimization method is based on
the mechanisms of natural selection [13]. Therefore, for fairly small variations around the
The optimal solution is sought from a population of operating point (2), the system model (1) can be
solutions using random processes. The search for the best approximated to the following linear delay system:
solution is made by creating a new generation of
solutions by successive application, to the current
population, of three operators: selection, crossing, and
mutation. These operations are repeated until a stop
criterion is reached [14]. We implement this technique on
the basis of the use of performance indices of integrated- (3)
absolute error (IAE), integrated-squared error (ISE),
integrated-time-absolute error (ITAE) and integrated-
time-squared error (ITSE)as objective functions that will
be minimized by GA, to determining the optimal PID
controller parameters so as to compensate the delay in
the system.
The system input corresponds to the probability of
packet loss, and the system output corresponds to the
II. System Model queue length of the router. By considering a negative
feedback control system [21], the linearized model can
This study revolves around the sharing of a be written as:
communication link between multiple senders at remote
locations and a single bottleneck (Fig. 1).
We adopt the fluid model in this study. The idea of (4)
these models consists in erasing a part of the
Copyright © 2017 Praise Worthy Prize S.r.l. - All rights reserved International Review of Automatic Control, Vol. 10, N. 1
34
Samira Chebli, Ahmed Elakkary, Nacer Sefiani
Often times, a crude model will suffice. In this case, a The results obtained in [23] – [25] when developing
first-order model which captures the overall behaviour of the stability condition above, show that we get separate
a system will do. Then we have: stabilizing regions of when sweeps three
different intervals (Figure 3).
(5)
where:
(6)
Copyright © 2017 Praise Worthy Prize S.r.l. - All rights reserved International Review of Automatic Control, Vol. 10, N. 1
35
Samira Chebli, Ahmed Elakkary, Nacer Sefiani
• there are a very large number of more or less good presented here was proposed by Fonseca et al. [34] and it
solutions is based on Pareto dominance which is an idea suggested
• there is no deterministic algorithm to calculate the best by Goldberg [13] to solve the problems introduced by
solution Schaffer [35]. The core of this approach is to use the
• the universe of the problem is little formalized Pareto optimality concept [36] to keep all the criteria
The flowchart in Fig. 4 presents a description of a intact, avoiding a priori comparison of the values of
standard GeneticinAlgorithm.
The flowchart figure 4 presents a description of a standard different criteria.
Genetic Algorithm:
The MOGA method was developed in [34] and [37].
Start In this method, the rank of an individual (the number of
an individual within others) is given by the number of
Ran individuals who dominate the individual.
dom Initial population Let’s consider an individual in the generation
ly
which is dominated by individuals. Then the rank of
Calculate Fitness the individual is:
YES
Rank 1 is assigned to all non-dominated individuals.
Satisfy
Optimization? The efficiency of an individual is calculated through
the following steps:
classify individuals according to rank,
NO
assign efficiency to an individual by performing an
interpolation from the best rank to the worst.
Select The algorithm describing the principle of the MOGA
method is given by:
Initialization of population
Crossover
Ran Evaluation of objective functions
dom
Assigning rank based on dominance
ly
Mutation Assignment of efficiency from rank
For to ( is the number of generations)
Random selection proportional to efficiency
Best Fitness
Crossing
Mutation
Evaluation of objective functions
Finish
Assigning rank based on dominance
Assignment of efficiency from rank
Fig.4. Steps of genetic algorithm evolution End for
Fig. 4. Steps of genetic algorithm evolution The principle of optimization by the MOGA approach
is presented in Figure 5. It is about searching the PID
Genetic algorithms are highlighted in comparison with
other classical methods of optimization by their ability to controller variables and among the stability
manipulate the coding of decision or control variables area.
which are chromosomes instead of functions or control
variables directly, and also because GAs work on a
population of individuals instead of a single point.
+
+
What makes this optimization technique more -
efficient is that it uses the values of the function to
optimize, whereas the other methods use auxiliary
information such as the derivative. As for the rules of
transitions of genetic algorithms, they are stochastic,
while most other methods have deterministic transition
rules.
Fig. 5. The optimization principle by genetic algorithm
V. Design of PID Controller Using From the closed loop of Figure 5 and review
Multiple Objective Genetic Algorithm established earlier, we choose some performance criteria
commonly used as objectives in time domain. The
(MOGA) mathematical description of these objectives is given by:
Genetic algorithms are used via different approach in The integral of the absolute value of the error signal
a variety of applications [30] – [33]. The MOGA method or IAE (Integral of the Absolute magnitude of the
Copyright © 2017 Praise Worthy Prize S.r.l. - All rights reserved International Review of Automatic Control, Vol. 10, N. 1
36
Samira Chebli, Ahmed Elakkary, Nacer Sefiani
Error):
(10)
(11)
Fig. 7. Unit step response for the different optimal PID controllers
By applying the Hermite-Biehler extension approach,
we find that the set of Stabilizing value is given In order to evaluate the dynamic behavior of the
below: system under a unit step response, we choose some
standard performance measure for their relevance to the
[ ] congested router system and the induced delay. Thus, the
selected criteria for qualifying and quantifying the
Figure 6 sketches a 3-dimensional representation of
performance of the system are:
the complete set of stabilizing PID controller variables
Stability: this is primary point of this brief, it is
and for the TCP/AQM system. From Figure 6,
evaluated by the settling time criterion.
it can be derived the stability range for each controller Speed: this is a point with an equally importance as
gain. So, the PID controller is stabilizing if and only if: the first one. Given the delay in the TCP router, the
[ ] challenge is therefore to minimize this dead-time. The
quality of this criterion is translated by the evaluation of
s the rise time and the settling time criteria.
Damping: it is characterized by the ratio between the
and: successive amplitudes of the oscillations of the output.
As these oscillations rapidly diminish, the system is
s damped. To characterize the quality of the damping we
Copyright © 2017 Praise Worthy Prize S.r.l. - All rights reserved International Review of Automatic Control, Vol. 10, N. 1
37
Samira Chebli, Ahmed Elakkary, Nacer Sefiani
measure the overshoot rate and the settling time values. choice of both the tuning approach of the PID controller
Table II summarizes these standard performances and the MOGA method for optimizing the PID
measures. parameters proposed in this brief.
TABLE II
NUMERICAL VALUES OF STANDARD PERFORMANCE MEASURES References
Criterion ISE IAE ITAE ITSE
[1] V. Misra, W. Gong, and D. Towsley. Stochastic differential
Rise Time (s) 0.3983 0.5589 0.6297 0.4611 equation modeling and analysis of TCP window size behavior.
Settling Time (s) 5.8501 3.1935 2.1873 3.7296 Technical Report ECE-TR-CCS-99-10-01, University of
Massachusetts, October 1999.
Overshot (%) 28.4493 10.6662 4.1763 21.9698 [2] H. I. Alshimaa, A. EL-Sayed& Z. Elsaghir, I. Z. Morsi,
―Enhanced Random Early Detection (ENRED)‖ International
Journal of Computer Applications (0975 – 8887) Vol. 92 – No.9,
As it is expected when using an optimization method April 2014.
based on Pareto optimality, one cannot improve one [3] V. Misra, W. Gong, and D Towsley. Fluid-based analysis of a
objective without deteriorating the others. As a result, we network of AQM routers supporting TCP flows with an
can see from Table II this conflict between the objectives application to RED. In ACM SIGCOMM, pages 151–160, August
2000.
that we opted for. The rise time provided when using the [4] H. S. Low, F. Paganini, and J.C. Doyle. Internet congestion
ISE objective function is ideally small, compared to the control. IEEE Control Systems Magazine, 22:28–43, Feb 2002.
one provided by the ITAE criterion. However, we obtain [5] D. Aawal and F. Granelli. Redesigning an active queue
a slightly higher overshoot rate by ISE criterion, but management system. In IEEE Globecoin. volume 2. pages 702—
706. December 2004.
ITAE gives a much better result. [6] C. V. Hollot and Y. Chait. Nonlinear stability analysis for a class
Overall, the values of both rise time and settling time of TCP/AQM networks. In IEEE Conference on Decision and
obtained via the various objectives used for the MOGA Control. pages 2309—23 14. December 2001.
optimization technique are quite small. This has a [7] 5. Manfredi. M. di Bernardo, and F. Garofalo. Robust output
feedback active queue management control in TCP networks. In
favorable impact on the behavior of the system, in IEEE Conference on Decision and Control. Pages 1004—1009.
particular on the reduction of the delay resulting from the December 2004.
congestion phenomenon. [8] K. B. Kim. Design of feedback controls supporting TCP based on
the state space approach. IEEE Trans. on Automat. control, 51(7),
July 2006.
[9] C. V. Hollot. V. Misra. D Towsley. and W. Gong. On desining
VII. Conclusion improved controllers for AQM routers supporting TCP flows. In
In this paper, we first studied the problem of IEEE INFOCOM. volume 3, pages 1726—1734. April 2001.
[10] C. V. Hollot. V. Misra. D Towsley. and W. Gong. Analysis and
stabilization of TCP networks by a delay systems design of controllers for AQM routers supporting TCP flows.
approach. It has been shown that the stabilization by the IEEE Trans. on Automat. Control. 47 :945—959. June 2002.
PID controller of this class of systems amounts to derive [11] G. J. Silva, A. Datta and S. P. Bhattacharyya, "PID controllers for
a necessary and sufficient condition for the roots of the time delay systems", Birkhäuser, MA, Boston, 2005.~
[12] B. Balachandran et al. (eds.), "Delay Differential Equations :
quasi-polynomial - the system characteristic equation - to Recent Advances and New Directions", Springer Science plus
have the real part negative. This was the subject of an Business Media LLC 2009.
extension of the Hermite-Biehler theorem whose proof DOI: 10.1007/978-0-387-85595-0 1
uses the property of interlacing roots of the quasi- [13] Goldberg, D. E. Genetic algorithm in search, optimization, and
machine learning, reading. MA: Addison Wesley, 1989.
polynomial [23] – [25]. Once the complete set of [14] Gao Ye, Zheng Tao: Improved genetic algorithms based on
stabilizing gains of the implemented PID controller to the chaotic mutation operation and its application. Multimedia
congested router is calculated via the previous approach, Technology (ICMT), 2010 International Conference on 1-3
then comes the step of selecting the optimal gains of the (2010).
[15] C.V. Hollot, V. Misra, D. Towsley, and W-B. Gong, A control
PID controller. This parameters optimization was carried theoretic analysis of red., Proceedings of IEEE INFOCOM
out by using the Multiple Objective Genetic Algorithm (2001), 10pp.
(MOGA) technique. This method combines the strengths [16] F. Kelly, A. Maulloo, and D. Tan, Rate control in communication
of both the genetic algorithm and the multi-objective networks : shadow prices, proportional fairness and stability,
Journal of the Operational Research Society 49 (1998), 237–252.
optimization translated by the ability of the first to work [17] K. B. Kim, Design of feedback controls supporting TCP based on
on a population of individuals instead of a single point modern control theories, IEEE Transactions on Automatic
and the use of the concept of Pareto optimality for the Control (2004).
second. To implement this optimization technique, four [18] K. B. Kim and S. H. Low, Design of receding horizon AQM in
stabilizing TCP with multiple links and heterogeneous delays,
performance criteria were chosen as objectives, namely Proc. of 4th Asian Control Conference (2002).
IAE, ISE, ITAE, and ITSE. [19] S. H. Low and D. E. Lapsley, Optimization flow control, i : basic
The dynamic performance of the delayed congestion algorithm and convergence, IEEE/ACM Transactions on
router was analyzed in terms of rise time, settling time Networking (1999), no. 7, 861–874, http: //netlab.caltech.edu.
[20] B. Miller, K. Avrachenkov, K. Stepanyan, and G. Miller, Flow
and overshoot rate. These three criteria characterize the control as stochastic optimal control problem with incomplete
quality of the stability, the speed, and the damping of the information, Proc. of IEEE Infocom (2004).
system. The results obtained for an example of [21] S. Chebli, A. Elakkary, N. Sefiani, N. Elalami, New PI
illustration taken from the literature reflect a good overall Stabilization of First Order Congestion Control of Active Queue
Management Routers, in Proceeding of Electrical and
performance of the system and thus reflect the judicious
Copyright © 2017 Praise Worthy Prize S.r.l. - All rights reserved International Review of Automatic Control, Vol. 10, N. 1
38
Samira Chebli, Ahmed Elakkary, Nacer Sefiani
Copyright © 2017 Praise Worthy Prize S.r.l. - All rights reserved International Review of Automatic Control, Vol. 10, N. 1
39