This action might not be possible to undo. Are you sure you want to continue?

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

DIPARTIMENTO DI INFORMATICA E SISTEMISTICA

**Nonlinear Model Predictive Control
**

Stability, Robustness and Applications

Davide M.

Raimondo

A Giovanni, Elena ed Alessandro

**A winner is a dreamer who never quits
**

N. Mandela

Giacomo Ferrario. Chiara Dalla Man and Dr. whose support was fundamental in these years of reserach. Dr. Last but not least. A particular thanks goes to all the members of the Articial Pancreas group. Matteo Rubagotti who had a crucial role in my work. Roberto Tessera. Andrea Facchinetti. A special thanks goes to Prof. Franceso Dinuzzo. Dr. First of all I would like to thank my supervisor. In these years he was not only a supervisor. Riccardo Porreca for the time spent together. I am really in debt to him for all the opportunities he gave me and all he teached me. for his teachings and his contribution to the development of my thesis. Eduardo Camacho. Steven Patek. he was a friend. Lalo Magni. Paola Soru. Daniel Limon for the greeting he gave me at University of Seville. and all the medical sta. Prof. Prof. Dr. Dr. in particular Dr. Dr. Prof. Antonio Ferramosca for his important help in the drafting of the thesis. Daniela Bruttomesso. Stefano Riverso. My gratitude goes to Prof. Dr. . Prof. Giuseppe De Nicolao. Prof. Claudio Cobelli. Riccardo Scattolini. Dr. A special thanks goes to Dr. Teodoro Alamo.4 Preface This thesis is the product of three years of research that I have carried out in the Department of Informatica and Sistemistica at the University of Pavia. Prof. thanks to my family and all the people that cares about me and that I care about. Marco Laganà. for his continuous support and encouragement. Boris Kovatchev. Luca Bossi. Claudio Vecchio and Dr. Marc Breton. in particular to Dr. Gilberto Pin. I would like also to thank Prof. Prof. Mircea Lazar. Thomas Parisini. Thanks to the colleagues Dr. that spent also some nights in hospital with me.

3 2.5 3. . . . . . . . . . . . . . . . . . . . . . . . . .2 Open-loop formulation . .4 Introduction . Regional Input-to-State Stability . . . . .6 3. . . . . . . . . . . . . . . . . . . . . . . 2. . . . . . Appendix . . . . . . . Problem statement . . . . . . . . . . . . . . .4. . . . 2. . . . . . . . . . . . Conclusions .1 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7 3. . . . . . . . . . . . . . . . . . . . . . 71 . . . . . . . . . . . . . . . . . . . . . . . . Closed-loop formulation . . . . . . . . . . . . .Contents 1 Introduction 2 Regional ISS for NMPC controllers 2. . . . . . . . . . . . . .3 3.2 3. . . . . . .6 2. Min-max model predictive control . 3 Min-Max NMPC: an overview on stability 3. . . . . . . . . . . . . . . . . . . . . .4. . . . .4 Introduction . . . . . . . . . 3. Appendix . . . . . . . . . .7 Remarks . . . . . . . . . . . . . . . .8 Stability . . .5 2. . . . . .1 Introduction . . 1 15 15 16 17 22 23 26 29 30 30 Nonlinear Model Predictive Control .2 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 3. . . . . . . . . . . . . 37 37 39 42 47 49 55 57 62 62 The auxiliary control law Example . . . . . . . . . .4. . . . . . . . . . . . . . Problem statement .1 2. . . . . . . . . . . . . . . . . . . . . Conclusions . . . 4 Min-Max NMPC: a min formulation with guaranteed robust stability 71 4.1 2. . . . . . . . Regional Input-to-State practical Stability . . . . . . . . . . . . . . . . . . . . .

Closed-loop formulation . . . . . . . . . . . Appendix . . . . . . . . . . .8 Problem statement . . . . . . . . . . . .6 Contents 4. . . . . . . . .2 5. . . . . . . . Simulation Results . .6 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 Decentralized NMPC: an ISS approach 6. . 6.4 Introduction . . . . . . . . . . . . . . . 6. . . .4. . . . . . . . . . . . . . . . . 7 Cooperative NMPC for Distributed Agents 131 . . . . . . . . . . .5 4. . . . . . . . Appendix . . . . . . .7 4. . . . . . . . . . . . . . . . . . . . .4 5. . . . . . . . . . . . . . . . 107 107 109 112 116 117 121 126 127 Nonlinear Model Predictive Control . . . .4. . . . .1 5. . . . . . . . . . . . . . . .3. . . . . .6 . . . Regional Input-to-State Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 5. .1 6. . 5 Robust Open-loop NMPC 5. . . . . . . . . . . . 87 87 89 92 93 94 95 96 99 99 Shrunk State Constraints Feasibility . . . . .5 5. . . . . . . . . . . .2 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 73 76 77 79 81 82 Example . . . . .2 6. . . . . . . . . . . . . . . . . . . . . . . .4 4. . . . . . . Guaranteed bound of the max function . . . . . . . . . . . .6 Conclusions . . .3 Introduction . . . Regional ISS for interconnected subsystems . . . . . . . . . . . . . . . . . . . Min-Max Nonlinear Model Predictive Control .5 6. . . . . . . . . . . . . . . . . . . . . . .3. . . . . .1 6. . Conclusions . . . . . . . . . . . . . . .3 6. . . . . . . . . . . Problem formulation . . . . .3 5. . . . . . . .3. . Appendix . . . . . . . Problem statement . . . . Proposed Min formulation: stability . .2 Open-loop formulation . . . . . . . . . . . . . . . . Robust MPC strategy 5. . Conclusions . . . .3 4. . . . . . . . . . . . . . . .2 4. . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6 Metrics to asses control performance Regulator design 8. . . . . . . . .4. . . . . . . . Conclusions . . . . . .2 Introduction . .3. . . . . Models for the Glucose-Insulin System . . . . . . . . .3 Regional ISS results . . . .5 8.3. . . . . . . .1 8. . . 8. . .6. . . . . . . . . . . . . . . .8 8. . 8. . . . . . .1 7. . . . . . . . .3 8. . 8 MPC of Glycaemia in Type 1 Diabetic Patients 8. Articial Pancreas 8. . 131 134 140 141 143 147 153 158 Stability properties of the single agents . . . . . . . . . . 8. . . . . . . . . . . . . . . . . . .6. . . . . . . . . . . . . . . . . . . . . . . . 8. .2 7. .3 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 8. . . . . . . . . . . . . . Problem statement . . . . . . . .Contents 7 7. . . . .1 8. . . . . . . . . . . . . . . . 175 175 177 180 181 185 185 197 201 201 210 212 213 215 216 Articial Pancreas Project . . . . . . . 7.1 Dalla Man et al. . . . . . . . . .5 Simulation results . . . .4 7. . . . . . . . . . . . . . . . . . . . . 9 Appendix of the Thesis Bibliography 221 227 . . . . . . .4 . .2 7. . . . . . . . . . . Results . . . . . .10 Conclusions . . . . . . . . . . model for the diabetic patient . . . . . . . . . . . . . . . 7. . . . . . . . . . . . . . . . . . . . . . . . . . .7 8. . . . .1 7. . . . . .3. . . . . . . . . . . . . . . . . . .2 . . .9 Clinical Scenario Control design procedure . . . . . . . . . . . Unconstrained LMPC NMPC . . . . . Glucose Regulation . . . . . . . Stability of the team of cooperating agents . . . . . . . . . . . . . . . Stability properties of the team of agents .2. . . .

.

Mayne et al. It is worth to note that this con- trol technique. MPC has proved enormously suc- cessful in industry mainly due to the ease with which constraints can be included in the controller formulation. The main items in the design of a predictive controller are: • • • the process model a performance index reecting the reference tracking error and the control action an optimization algorithm to compute a sequence of future control . Rossiter 2003. the books [Maciejowski 2002. not only in the process industry but also other processes ranging from automotive to clinical anaesthesia. Rawlings 2000. Camacho & Bordons 2004] and the survey papers [Morari & H. Findeisen et al. 2003.Chapter 1 Introduction Model Predictive Control (MPC). refers to a class of algorithms which make explicit use of a process model to optimize the future predicted behavior of a plant. has achieved great popularity in spite of the original lack of theoretical results concerning some crucial points such as stability and robustness. Qin & Badgwell 2003. In fact. Magni & Scattolini 2007]). Predictive control uses a model of the system to obtain an estimate (prediction) of its future behavior. Several recent publications provide a good introduction to theoretical and practical issues associated with MPC technology (see e. Originally developed to cope with the control needs of power plants and petroleum reneries. One of the reasons for the success of MPC algorithms consists in the intuitive way of addressing the control problem.g. Lee 1999. During the past 30 years. 2000. also known as moving horizon control or receding horizon control. a theoretical basis for this technique started to emerge more than 15 years after it appeared in industry. it is currently successfully used in a wide range of applications.

a nite horizon optimal control problem is solved over a prediction horizon N . At each sampling time t. Introduction signals that minimizes the performance index subject to a given set of constraints • the receding horizon strategy. using the current state x of the process as The optimization yields an the initial state. constraints and control objectives. The on-line optimization problem takes account of system dynamics.2 Chapter 1. optimal control sequence. Past Future ymax r Setpoint Measured Outputs Predicted Outputs ymin Prediction horizon umax Closed-loop inputs Control Horizon Open-loop inputs umin -4 -3 -2 -1 k +1 +2 +3 +4 +5 +6 +7 +8 +9 Sampling time Figure 1. Figure 1. At the next time instant the horizon is shifted one sample and the optimization is restarted with the information of the new measurements. according to which only the rst element of the optimal control sequence is applied on-line.1 depicts the basic principle of model predictive control.1: A graphical illustration of Model Predictive Control . and only the control action for the current time is applied while the rest of the calculated sequence is discarded.

In this case. 2005. will provide quite better results. Lee 1999] and the reference therein). the use of a nonlinear model implies higher diculties in the calculation of the control law and in the stability analysis of the obtained closed-loop system. dual-mode nonlinear MPC [Michalska & Mayne 1993]. model/plant mismatch or unknown disturbances are always present. By now. i. Clearly the use of a more accurate model. Rawlings 2000. the maintenance of certain properties such as stability and performance in the presence of uncertainty. linear and nonlinear MPC are distinguished. the results obtained by using a LMPC are poor in term of performance and often not sucient in order to cope with the process requirements. De Nicolao et al. Linear MPC refers to a family of MPC schemes in which linear models are used to predict the system dynamics. 2001a]. dictive tems: control schemes MPC with with By now there are several prestability terminal for nonlinear sys- guaranteed zero state nonlinear equality constraint [Chen & Shaw 1982. On the other hand. linear MPC theory is fairly mature (see [Morari & H.3 In general. Mayne & Michalska 1990]. nonlinear MPC with a suitable terminal state penalty and terminal inequality state constraint [Parisini & Zoppoli 1995. Many theoretical and prac- tical issues have been addressed [Allgöwer et al. 2000].e. Chen & Allgöwer 1998. Model predictive control for nonlinear systems (NMPC) has received considerable attention over the past years. 1999.e. Qin & Badgwell 2003]. Linear MPC approaches have found success- ful applications. discrete-time systems (when a terminal equality constraint is employed). In practical applications. constrained. In [Keerthi & Gilbert 1988. Grimm et al. especially in the process industries [Qin & Badgwell 2000. Mayne & Michalska 1990] the value function (of a nite horizon optimal control problem) was rst employed as a Lyapunov function for establishing stability of model predictive control of time-varying. However many systems are inherently nonlinear. the assumption that system is identical to the model used for prediction is unrealistic. a nonlinear model. Mayne et al. The introduction of uncertainty in the system description raises the question of robustness. . Nowadays the value function is universally employed as a Lyapunov function for stability analysis of MPC. De Nicolao et al. i. nonlinear. 1998b. Keerthi & Gilbert 1988. 2000. and nonlinear MPC without terminal constraint [Jadbabaie & Hauser 2001. Magni et al. Limon et al. 2006b]. In fact.

it is well known that asymptotic stability of the controlled system does not suce to ensure robustness. In the case of constrained MPC. [Sontag & Wang 1996]. 1994]. Continuity of the optimal cost is a property dicult to be proved in general. a nominal stabilizing MPC may exhibit zero-robustness. see e. 2000]. [Huang et al. [Angeli et al. [Gao & Lin 2000]. 2004]. 2002b]. De Nicolao et al. In [Meadows et al. The concept of ISS was rst introduced in [Sontag 1989] and then further exploited by many authors in view of its equivalent characterization in terms of robust stability. When the plant is unconstrained. 1995]: the trajectories (depicted with triangles and daggers respectively) obtained starting from two close initial states (depicted inside a circle) are com- .2 shows this fact by recalling the example of [Meadows et al. Recently [Limon et al. are considered). 1998a]. In [De Nicolao et al. [Nesi¢ & Laila 2002]. the authors demonstrate that Lipschitz continuity of the control law provides some robustness under decaying uncertainties. [Jiang & Wang 2001]. [Kellett & Teel 2004] and [Grimm et al. 1997]. Introduction Input-to-State Stability (ISS) is one of the most important tools to study the dependence of state trajectories of nonlinear continuous and discrete time systems on the magnitude of inputs.4 Chapter 1. Inherent robustness of unconstrained MPC can also be derived from the inverse optimality of the controller [Magni & Sepulchre 1997]. [Jiang et al. Now. 2005]. which can represent control variables or disturbances. Then further conditions must be considered. several variants of ISS equivalent to the original one have been developed and applied in dierent contexts (see e. the resulting MPC is known to be robust under certain variations of the gain of inputs [Glad 1987. 2005]). 2004] have demonstrated that continuity of the optimal cost function plays an important role in the robustness of nominal MPC. [Sontag & Wang 1995a]. In MPC. 1996].g. This is due to the fact that the feedback law comes from the solution of a constrained optimization problem (when constraints. similar results are obtained from the continuity of the optimal cost function. in [Scokaert et al. 1995] it was shown that MPC could generate discontinuous feedback control law with respect to the state variable x even if the dynamic system is continuous.g. dissipativity and input-output stability. [Huang et al. Geromel & Da Cruz 1987. Figure 1. As studied in [Grimm et al. as for example the terminal constraint.

rstly. When the plant is unconstrained and the terminal constraint is not active [Jadbabaie & Hauser 2001]. This further requirement adds complex- Several results have been proposed concern[Mayne et al. Note that in the example just the terminal constraint is present. Another relevant case is provided for linear systems with polytopic constraints in [Grimm et al. namely the robust constraint satisfaction. or when only constraints on the inputs are present [Limon et al. satisfaction of the constrained for any possible uncertainty. 2006b]. ity to the MPC synthesis. To this aim. this problem remains open [Lazar 2006]. Magni & Scattolini 2007] . must be ensured. discontinuity of the control law is avoided. 2000. ing robust MPC (see e. Therefore. MPC control law could be a priori discontinuous and Figure 1. When the model function is discontinuous (as for instance the case of hybrid systems) these results can not be used and then.g. 2004]. 1995] consequently also the closed-loop system dynamics and the value function. Robust MPC stems from the consideration of the uncertainties explicitly in the design of the controller.5 pletely dierent. This could be a problem since continuity of the value function is required in most of the literature in order to prove robust stability of MPC (see [Mayne et al.2: Closed-loop state trajectories of example proposed in [Meadows et al. 2000]).

6

Chapter 1. Introduction

for

a

survey,

the

reference

therein,

and

some

of

the

most

re-

cent papers [Grimm et al. 2004, Grimm et al. 2007, Mhaskar et al. 2005, DeHaan & Guay 2007, Limon et al. 2005, Rakovi¢ et al. 2006b]). In particular two dierent approaches have been followed so far to derive robust MPC algorithms:

• •

open-loop MPC formulations with restricted constraints, see for example [Limon et al. 2002a], [Grimm et al. 2003] min-max ple open and closed-loop formulations, see for exam-

[Chen et al. 1998,

Magni et al. 2001b,

Gyurkovics 2002,

Magni et al. 2003, Magni & Scattolini 2005, Limon et al. 2006a]. The rst method for the design of robust MPC consists in minimizing a nominal performance index while imposing the fulllment of constraints for each admissible disturbance. This calls for the inclusion in the problem formulation of tighter state and terminal constraints. The idea was introduced in [Chisci et al. 2001] for linear systems and applied to nonlinear systems in [Limon et al. 2002a] [Raimondo & Magni 2006]. The main drawback of this open-loop strategy is the large spread of trajectories along the optimization horizon due to the eect of the disturbances and leads to very conservative solutions or even to unfeasible problems. With a signicant increase of the computational burden, an alternative approach is the second one, that consists in solving a min-max optimization receding open-loop problem, originally in proposed in the context of robust in an re-

horizon

control

[Witsenhausen 1968]. performance index is

Specically, minimized

formulation,

the

with

spect to the control sequence for the worst case, i.e.

the disturbance

sequence over the optimization horizon which maximizes the performance index. However, this solution may still be unsatisfactory, since the

minimization with respect to a single control prole may produce a very little domain of attraction and a poor performance of the closedloop systems. As demonstrated in [Chen et al. 1997, Bemporad 1998]

if some feedback is added by pre-stabilizing the plant, then the conservativeness is reduced. Recently, a closed-loop formulation of the

min-max controller has been proposed to reduce this conservativeness [Magni et al. 2001b, Scokaert & Mayne 1998, Mayne et al. 2000]. predictive control technique, a vector of feedback control In this is

policies

7

considered in the minimization of the cost in the worst disturbance case. This allows to take into account the reaction to the eect of the uncertainty in the predictions at expense of a practically untractable optimization problem. In this context robust stability issues have been recently studied and some novel contributions on this topic have appeared in the literature [Scokaert & Mayne 1998, Gyurkovics 2002, Kerrigan & Mayne 2002, Magni et al. 2003, Löfberg 2003, Gyurkovics & Takacs 2003, Fontes & Magni 2003, Magni & Scattolini 2005,

Kerrigan & Maciejowski 2004,

Lazar 2006, Limon et al. 2006a, Lazar et al. 2008]. The ISS property has been recently systems introduced controlled in the study of

nonlinear for

perturbed

discrete-time

with

MPC

(see

example

[Goulart et al. 2006,

Limon et al. 2006a,

Kim et al. 2006,

Grimm et al. 2007, Lazar et al. 2008]), Note that, in order to apply the ISS property to MPC closed-loop systems global results are in general not useful due to the presence of state and input constraints. On the other hand local results, see e.g. [Jiang & Wang 2001, Jiang et al. 2004], do not allow the

analysis of the properties of the predictive control law in terms of its region of attraction.

Thesis overview

In this thesis, the ISS tool is used in order to study the stability properties of nonlinear perturbed discrete-time systems controlled with MPC. As previously stated, since global and local results do not allow the analysis of the properties of MPC, the notion of regional ISS will be introduced. The equivalence between the ISS property and the existence of a suitable possibly discontinuous ISS-Lyapunov function will be established. As discussed before, since MPC could provide a discontinuous control law with respect to the state

x,

the value function, used as ISS-Lyapunov function, should

be possible discontinuous. Just the continuity of the Lyapunov function at the equilibrium point is required. The developed tool will be used in order to analyze and synthesize robust MPC controllers. Both open-loop and min-max robust MPC are considered. In particular it will be shown that open-loop MPC algorithms can guarantee ISS of the closed-loop systems while min-max MPC algorithms with standard stage cost can only guarantee Input-to-State practical Stability (ISpS). Dierent stage cost and dual-mode strategy will be used in

8

Chapter 1. Introduction

order to recover the ISS for closed-loop system. Moreover, since the minmax problem is very computational demanding, a relaxed formulation, in which the max stage is replaced by a simple suitable choice of an uncertain realization, will be presented. Then, in order to eciently consider state

dependent disturbances and improve existent algorithms, a new open-loop robust MPC algorithm will be presented. Subsequently the problems of decentralized and cooperative NMPC control will be addressed and stability properties will be stated by using the concepts of ISS and small-gain theorem. Finally, the problem of glucose control in type 1 diabetic patients will be coped with by using linear and nonlinear model predictive controllers.

Thesis structure

Chapter 2: Regional ISS for NMPC controllers.

bances. In this chapter, regional input-to-state stability is introduced and studied in order to analyze the domain of attraction of nonlinear constrained systems with disturISS is derived by means of a possible non smooth ISS-Lyapunov

function with an upper bound guaranteed only in a sub-region of the domain of attraction. These results are used to study the ISS properties of nonlinear model predictive control algorithms. Chapter 2 contains results published in:

•

[Magni et al. 2006b]: L. Magni, D. M. Raimondo and R. Scattolini. Input-to-State Stability for Nonlinear Model Predictive Control. In Proceedings of 45th IEEE CDC, pages 48364841, 2006.

•

[Magni et al. 2006a]: L. Magni, D. M. Raimondo and R. Scattolini. Regional Input-to-State Stability for Nonlinear Model Predictive Control. IEEE Transactions on Automatic Control, vol. 51, no. 9, pages 15481553, 2006.

**Chapter 3: Min-Max NMPC: an overview on stability.
**

of uncertain nonlinear systems subject to constraints.

Min-

Max MPC is one of the few techniques suitable for robust stabilization Stability issues as

well as robustness have been recently studied and some novel contributions on this topic have appeared in the literature. In this chapter, a general

9

framework for synthesizing min-max MPC schemes with an a priori robust stability guarantee is distilled. Firstly, a general prediction model that

covers a wide class of uncertainties, which includes bounded disturbances as well as state and input dependent disturbances (uncertainties) is introduced. Secondly, the notion of regional Input-to-State Stability (ISS) is extended in order to t the considered class of uncertainties. Then, it is established that only the standard min-max approach can only guarantee practical stability. Two dierent solutions for solving this problem are proposed. The rst one is based on a particular design of the stage cost of the performance index, which leads to a

H∞

strategy, while the second one is based on a dual-mode

strategy. Under fairly mild assumptions both controllers guarantee Inputto-State Stability of the resulting closed-loop system. Moreover, it is shown that the nonlinear auxiliary control law introduced in [Magni et al. 2003] to solve the

H∞ problem can be used, for nonlinear systems ane in control, in

all the proposed min-max schemes and also in presence of state independent disturbances. A simulation example illustrates the techniques surveyed in this article. The results presented in Chapter 3 are published in:

•

[Raimondo et al. 2007b] D. M. Raimondo, D. Limón, M. Lazar, L. Magni and E. F. Camacho. Regional input-to-state stability of minmax model predictive control. In Proceedings of IFAC Symp. on

Nonlinear Control Systems, 2007.

•

[Raimondo et al. 2008b]: D. M. Raimondo, D. Limón, M. Lazar, L. Magni and E. F. Camacho. nonlinear systems: Min-max model predictive control of To appear in

A unifying overview on stability.

European Journal of Control, 2008.

**Chapter 4: Min-Max Nonlinear Model Predictive Control: a min formulation with guaranteed robust stability.
**

This chapter presents a relaxed formulation of the min-max MPC for constrained nonlinear systems. In the proposed solution, the maximization problem is replaced by the simple evaluation of an appropriate sequence of disturbances. This reduces dramatically the computational burden of the optimization problem and produces a solution that does not dier much from the one obtained with the original min-max problem. Moreover, the

10

Chapter 1. Introduction

proposed predictive control inherits the convergence and the domain of attraction of the standard min-max strategy. The results of Chapter 4 are presented in:

•

[Raimondo et al. 2007a]: D. M. Raimondo, T. Alamo, D. Limón and E. F. Camacho. Towards the practical implementation of min-max In Proceedings of 46th IEEE

nonlinear Model Predictive Control. CDC, pages 12571262, 2007.

•

[Raimondo et al. 2008a] D. M. Raimondo, T. Alamo, D. Limón and E. F. Camacho. Min-Max Nonlinear Model Predictive Control: a Min formulation with guaranteed robust stability. Submitted to a journal, 2008.

**Chapter 5 Robust MPC of Nonlinear Systems with Bounded and State-Dependent Uncertainties.
**

In this chapter, a robust model predictive control scheme for constrained discrete-time nonlinear systems aected by bounded disturbances and statedependent uncertainties is presented. Two main ideas are used in order to improve the performance and reduce the conservatism of some existing robust open-loop algorithms. In order to guarantee the robust satisfaction

of the state constraints, restricted constraint sets are introduced in the optimization problem, by exploiting the state-dependent nature of the considered class of uncertainties. Moreover, dierent control and prediction

horizons are used. Unlike the nominal model predictive control algorithm, a stabilizing state constraint is imposed at the end of the control horizon in place of the usual terminal constraint posed at the end prediction horizon. The regional input-to-state stability of the closed-loop system is analyzed. A simulation example shows the eectiveness of the proposed approach. Chapter 5 is based on:

•

[Pin et al. 2008a]: G. Pin, L. Magni, T. Parisini and D. M. Raimondo. Robust Receding - Horizon Control of Nonlinear Systems with State Dependent Uncertainties: an Input-to-State Stability Approach. In

Proceedings of American Control Conference, Seattle, Washington, USA, June 11-13, pages 1667-1672, 2008.

•

[Pin et al. 2008b] G. Pin, D. M. Raimondo, L. Magni and T. Parisini. Robust Model Predictive Control of Nonlinear Systems with Bounded

Provisionally accepted for publication in IEEE Transaction on Automatic Control. A decentralized MPC algorithm for nonlinear systems. The cooperative control problem is formulated in a receding-horizon framework. which operate in a common environment and exchange-delayed information between them. Chapter 7 Cooperative NMPC for Distributed Agents. A decentralized MPC algorithm for nonlinear systems. based on locally computed control laws. where the control laws depend on the local state variables (feedback action) and on delayed information gathered from cooperating neighboring agents (feedforward action). 17. • [Raimondo et al. 2007d]: D. 2007c] D. In Proceedings of IFAC Symp. Magni and R. Magni and R. Scattolini. 2007. Decentralized MPC of nonlinear systems: An input-to-state stability approach. on Nonlinear Control Systems. A rigorous stability analysis exploiting the input-to-state . each one locally controlled with an MPC algorithm guaranteeing the ISS property. Then. • [Raimondo et al. 2007. 2008. M. M. The overall system under control is composed by a number of subsystems. which are computed by exchanging delayed state information with a subset of neighboring agents. the main stability result is derived by considering the eect of interconnections as perturbation terms and by showing that also the overall system is ISS. International Journal of Robust and Nonlinear Control. Scattolini. Each agent is assumed to evolve in discrete-time. Raimondo L. Raimondo L. Chapter 6 Decentralized NMPC: an ISS approach. M.11 and State-Dependent Uncertainties. on Nonlinear Control Systems. with decoupled nonlinear discrete-time dynamics. 2007e] D. The results in Chapter 6 are appeared in: • [Raimondo et al. Scattolini. pages 1651-1667. Raimondo L. Magni and R. 2007. This chapter addresses the problem of cooperative control of a team of distributed agents. vol. Both open-loop and closed-loop min-max formulations of robust MPC are considered. This chapter presents stabilizing decentralized model predictive control algorithms for discrete-time nonlinear systems. In Proceedings of IFAC Symp.

An in silico trial consisting of 100 patients is used to assess the per- formances of a linear and a nonlinear state feedback model predictive controller designed on the basis of the in-silico model. L. no. Moreover. vol. The results in Chapter 7 are published in: • [Franco et al. one of the main advantages of a nonlinear approach is the possibility to use a nonlinear cost function based on the risk index dened in [Kovatchev et al. 2007a]: C. 53. Raimondo. M. on Automatic Control. Cooperative Constrained Control of Distributed Agents With Nonlinear Dynamics and Delayed Information Exchange: A Stabilizing Receding-Horizon Approach. Chapter 8 MPC of Glycaemia in Type 1 Diabetic Patients. Magni. experiments using the developed linear model predictive controller are in progress at the Charlottesville and Padova hospitals. Franco. C. IEEE Trans. Raimondo. Rizza. In this chapter. Kovatchev and C. GIM. Cobelli. vol. Dalla Man. A recently developed in-silico model of glucose metabolism is employed to generate virtual patients on which control algorithms can be validated against interindividual variability. no. • [Magni et al. 1. 2007. Model Predictive . pages 324328. pages 323330. Polycar- pou and D. 2. The obtained results encourage a deeper investi- gation along this direction. C. M. 2005]. It is worth to note that currently. M. Parisini. 2008. D. B.12 Chapter 1. 2008]: E. D. G. 2007b]: L. 4. Raimondo. T. The stability of the team of agents is then proved by utilizing small-gain theorem results. De Nicolao. Cobelli. L. Introduction stability properties of the receding-horizon local control laws is carried out. The simulation experiments highlight the increased eectiveness of the meal announcement signal with respect to the linear MPC due to a more accurate nonlinear model. the feedback control of glucose concentration in type 1 diabetic patients using subcutaneous insulin delivery and subcutaneous continuous glucose monitoring is considered. Bossi. Magni. Journal of Diabetes Science and Technology. Dalla Man. Chapter 8 contains results based on: • [Dalla Man et al. R. Simulation Software of Meal Glucose-Insulin Model. M. A.

S. S. Breton. Moreover. M. 2008. G. Cobelli. Kovatchev. Provisional Application Patent. 60/984. M. Raimondo. Dalla Man. Cobelli. 2007a]: L. G. Seoul. Magni. Each chapter has an appendix that contains the proofs of all lemmas and theorems. pages 630635. 2008a]: L. D. the notations and basic denitions used in the thesis are are gathered in the Chapter 9. 2008. De Nicolao. M. C. pages 804812. Kovatchev. de Nicolao. 6. 2008]: B. Raimondo. • [Magni et al. Patek. 11/02/2007. 2007. Cobelli and B. Evaluating the ecacy of closed-loop glucose regulation via controlvariability grid analysis (CVGA). Journal of Diabetes Science and Technology. Korea. Journal of Diabetes Science and Technology. July 6-11. • [Magni et al. Appendix of the Thesis. D. Kovatchev and C. no. Filed • [Magni et al. In Proceedings of 17th IFAC World Congress. no.S. M. D. U. In Proceedings of 17th IFAC World Congress. vol. B. 2008b]: L. Seoul. Raimondo. Dalla Man.956. G. C. pages 4234-4239. Raimondo. Breton. 4. Model Predictive Control of glucose concentration in subjects with type 1 diabetes: an in silico trial. Magni. Korea. July 6-11. In Silico Testing and in Vivo Experiments with Closed-Loop Control of Blood Glucose in Diabetes. M. pages 42464251. 1. C. • [Kovatchev et al. M. 2008. De Nicolao. 2. vol. Serial No. D. Patek and C. Magni.13 Control of Type 1 Diabetes: An in Silico Trial. . Model Predictive Control Based Method for Closed-Loop Control of Insulin Delivery in Diabetes Using Continuous Glucose Sensing.

.

. and the equivalence between the ISS property and the existence of a suitable Lyapunov function is established. . .4.1 Introduction Input-to-state stability (ISS) is one of the most important tools to study the dependence of state trajectories of nonlinear continuous and discrete time systems on the magnitude of inputs. . . . . . .5 2. . . . . . . . . . . global results are in general not useful in view of the presence of state and input constraints. in this chapter. this . . . . . . . . . . . . 15 16 17 22 Problem statement Regional Input-to-State Stability . 2. . . .2 2. Closed-loop formulation . . . . local results. . .2 2. . . . . . which can represent control variables or disturbances. . . . . do not allow to analyze the properties of the predictive control law in terms of its region of attraction. . . . . . . . . . . . . . . . . . . . .4 Introduction . . . . . . . . . . . . .7 Open-loop formulation . . On the other hand. . . . . . . . . . . . . . . . . . Nonlinear Model Predictive Control . . . . . .6 2. . Conclusions . Appendix . . . . . . . . . . . 2. . .3 2. . . . . . . . . . . . Notably.4. . . . . . . . . . . . . . Then. see also [Nesi¢ & Laila 2002]. . . . . . . the notion of regional ISS is initially introduced. . .1 2. . . .Chapter 2 Regional ISS for NMPC controllers Contents 2. . .1 2. . . 23 26 29 30 30 Remarks . . . . . . . . . In order to apply the ISS property to MPC.

w is such that w∈W where (2. uk ∈ IRm is the current control vector. For simplicity of notation. uk ) + wk . where state and terminal constraints are modied to improve robustness. Regional ISS for NMPC controllers Lyapunov function is not required to be smooth nor to be upper bounded in the whole region of attraction. The achieved results are used to derive the ISS properties of two families of MPC algorithms for nonlinear systems. such that f (xk .16 Chapter 2.2 Problem statement Assume that the plant to be controlled is described by discrete-time nonlinear dynamic model xk+1 = F(xk . 2002a]. dk ∈ IRq is the disturbance term. let nominal model.1).3) W is a compact set containing the origin. 2003]. k ≥ 0 where (2. An estimation of the region where the state of the system converges asymptotically is also given.1 1. dk ). uk ) denote the xk+1 = f (xk . 2. k ≥ 0 where (2. which indeed are dicult and not immediately veriable hypothesis. uk . The disturbance f (0. 2. i. uk ) denote the additive uncertainty. The system is supposed to fulll the following assumption. The rst one relies on an openloop formulation for the nominal system.e.1) xk ∈ IRn is the state. uk . Given the system (2. . 0) = 0.2) wk F(xk . see also [Limon et al. No continuity assumptions are re- quired on the value function or on the resulting MPC control law. dk ) − f (xk . with W sup known. The second algorithm resorts to the closed-loop min-max formulation already proposed in [Magni et al. Assumption 2. it is assumed that the origin is an equilibrium point.

e. i.5) X and U are compact sets. there exists a positive constant Lf such that (2.3. In the following section it is presented a suitable framework for the analysis of stability: the regional ISS. wk ). w).3 Regional Input-to-State Stability Consider a discrete-time autonomous nonlinear dynamic system described by xk+1 = F (xk . b ∈ X and all u ∈ U. both containing the origin as an interior point. Regional Input-to-State Stability 17 3. x. The state of the plant xk The control objective consists in designing a control law u = κ(x) such that it steers the system to (a neighborhood of ) the origin fullling the constraints on the input and the state along the system evolution for any possible disturbance and yielding an optimal closed-loop performance according to certain performance index. The state and the control variables are restricted to fulll the following constraints x∈X u∈U where (2.4) (2. . k ≥ 0. 2.6) |f (a. u)| ≤ Lf |a − b| for all a. can be measured at each sample time. The transient of the system (2. where (2. 4. xk ∈ IRn is the state and wk ∈ IRq is an unknown disturbance. 5. u) − f (b.2. The map f : IRn × IRm → IRn is Lipschitz in x in the domain X × U.7) with initial state x0 = x and disturbance sequence w ¯ is denoted by x(k. This system is supposed to fulll the following ¯ assumptions.7) F : IRn × IRq → IRn is a nonlinear possibly discontinuous function.

x.18 Assumption 2. if Ξ is a RPI set for system (2. The origin of the system is an equilibrium point.3 x=0 ¯ and w=0 with respect to disturbances and initial conditions. the system (2. w ∈ MW satises the Denition 2. Let introduce the following denitions. w)| ≤ γ( w ) + ε ¯ for all (2.7) with (Input-to-State Stable) in Ξ ⊂ Rn including the w ∈ MW . k) + γ( w ). ∃T = T (ε. Ξ ⊂ Rn including the origin as an interior point. ∀k ≥ 0 ¯ for all |¯| ≤ δ x and all |wk | ≤ δ . there exists a δ>0 such that (2. x and all k ≥ T. Assumption 2.1 (UAG in Ξ) |x(k. i.e. Given a compact set Denition 2. is said to be (2.7) and if there exists a K-function γ such that for each ε > 0 and ν > 0. ν ) such that Given a compact set Denition 2. 0) = 0. 2. Regional ISS for NMPC controllers 1.9) x∈Ξ ¯ with |¯| ≤ ν .3 (ISS in Ξ) ISS there exist a origin as an interior point. for each ε > 0. x.8) W is a compact set containing the origin.7) with w ∈ MW satises the U AG (Uniform Asymptotic Gain) property in Ξ.10) |x(k.11) .7) and if Ξ if KL-function β and a Ξ is a RPI set for system K-function γ such that |x(k. The disturbance F (0.2 (LS) The system (2.2 Chapter 2.7) is continuous at W sup known. x. w)| ≤ β(|¯|. with The solution of (2. ∀k ≥ 0. ¯ x x (2.7) with LS (Local Stability) property if. ∀¯ ∈ Ξ. the system (2. w)| ≤ ε. w is such that w∈W where (2.

2 and 2.3 hold. because the continuity of function F (·. s > 0 and time T = T (ε. ·) is not used. it was shown that. Note that. Regional Input-to-State Stability ISS 19 Note that. ∀k ≥ 0.1 if and only if Suppose that Assumption 2. Ξ implies LS . ·) continuously dierentiable). the following lemma can be stated. the ISS property is equivalent to the conjunction of UAG and LS. this is always satised. The proof is equal to the one given in [Sontag & Wang 1995a] and [Gao & Lin 2000].1 UAG in Consider system (2. if a RPI compact set Ξ is considered. there is an L > 0 such that |x(t.2. System (2. s) > 0 as in Denition 2. for continuous-time and discrete-time case respectively. by causality. But if Ξ is a ¯ x RPI compact set. Lemma 2. Proposition 2.3. If Assumptions 2. 1996].7) is Ξ if and only if U AG in Ξ holds. ∀¯ ∈ Ξ.2 and 2. Lemma 2.7 in [Sontag & Wang 1995a] carefully.3. since all the possible trajectories are bounded. By examinating the proof of Lemma 2. for continuous-time and discrete-time systems respectively (with F (·. x. it turns out that LS property is redundant. in the proof.3 is necessary in order to have ISS. r. all |¯| ≤ s and all ||w|| ≤ r.3 holds. one can see that it also applies to discontinuous systems (both continuous and discrete-time). Hence. the continuity is necessary just in or- der to use Proposition 5. If also Assumption 2. if the solution of (2.2 ISS in Suppose that Assumptions 2. the following proposition holds.7) is not continuous at x = 0 ¯ and . w)| ≤ L for all 0 ≤ t ≤ T .7) is ISS in Ξ U AG in Ξ and LS hold. by Denition 2.3 hold.7). In fact. x. ¯ x x Recall that (2.11) by in Ξ would result if one |x(k. w)| ≤ β(|¯|.2 holds. In fact. the same denition of would replace (2. System (2. k) + γ( w[k−1] ). In [Sontag & Wang 1995a] and in [Gao & Lin 2000].1.12) w[k−1] denotes the truncation of w at k − 1. In fact. Assumption 2. to prove that given ε.1 in [Lin et al.

one has to verify that Θ {x : V (x) ≤ b(W sup )} ⊆ IΩ (2. Regional ISS for NMPC controllers w = 0 hold. Ξ is a compact RPI set including the origin as an interior point 2.1.14) K∞ -function α3 and a K-function σ such that (2.20 Chapter 2. there exist a suitable (2.13) (2.7).15) ∆V (x) for all V (F (x. In order to clarify the relation between the sets introduced in the denition see Figure 2.4 (ISS-Lyapunov function in Ξ) A function V : IRn → IR≥0 is called an ISS-Lyapunov function in Ξ for system (2. α2 such that interior V (x) ≥ α1 (|x|). there exist a suitable a K∞ -function) and a suitable constant ρ such that (id − ρ) is cθ > 0. given a set an disturbance sequence Θw ⊆ IΩ w ∈ MW . and a pair of Ω ⊆ Ξ (including the origin as an suitable K∞ -functions α1 . Note that. w∈W K∞ -function ρ (with 4. Remark 2. there exist a compact set point).17) . such that. by (2. |x|δΩ > cθ } (including the origin as interior point) dened as follows Θw where {x : V (x) ≤ b(||w||)} with (2. there exists a nonempty compact {x : x ∈ Ω.16) b −1 α4 ◦ ρ−1 ◦ σ. ∀x ∈ Ω 3. function V is continuous at the origin. ISS does not The ISS property is now related to the existence of a suitable a-priori non smooth Lyapunov function dened as follows. ∀x ∈ Ξ V (x) ≤ α2 (|x|).1 Note that.14). with respect to disturbances and initial conditions. in order to verify that Θw ⊆ IΩ for all w ∈ MW . α4 −1 α3 ◦ α2 . if 1.13) and (2. w)) − V (x) ≤ −α3 (|x|) + σ(|w|) and all x ∈ Ξ. Denition 2.

in view of (2. as for example in the MPC. Theorem 2. Note that.3 hold. Hence. in order to verify that point 4 of Denition 2. of Θ) .1 Suppose that Assumptions 2. considering also Remark ¯ Θ −1 {x : |x| ≤ α1 ◦ b(W sup )} ⊆ IΩ (2.1: Example of sets satisfying Denition 2. This region depends on the bound on ¯ Θ ⊇ Θ. for all Remark 2.7) can be stated. In some cases.2.13) and Remark 6.1 gives an estimation of the region Θw where the state of the system converges asymptotically. If system (2.4 is satised. as well as on α2 .1. one has to verify that V is not known in explicit form.4 Now. Regional Input-to-State Stability 21 Figure 2. then it is ISS in Ξ w ∈ MW .18) In fact. w)|Θw = 0. since the size of ¯ Θ (as well as the size α1 .2 Theorem 2. limk→∞ |x(k. α3 and ρ. the following sucient condition for regional ISS of system (2.3.2 and 2. ¯ and. the function 6. w through σ.7) admits an ISS-Lyapunov function in disturbance sequences Ξ.1. x.

2: Relation between sets 2. In order to clarify the relation between the sets. the regularities of the value function or of the resulting control law are not assumed. α2 is directly related to an accurate estimation of the upper bound is useful in order to compute a smaller region of attraction Θ. 2005] that copes also discontinuous Lyapunov function).4 Nonlinear Model Predictive Control In this section. Moreover. sup = 0. Regional ISS for NMPC controllers α2 . since not necessary.2) fulll the ISS property. In the following. .22 Chapter 2.1 are used to analyze under which assumptions open-loop and closed-loop min-max formulations of stabilizing MPC for system (2. see Figure 2. Figure 2. the results derived in Theorem 2.2. then σ(W sup ) = 0 so that asymptotic stability is guaranteed if W (for this particular case see [Lazar et al.

ˆ with xt = x ˆ ¯ 3. Denition 2.t+N −1|t] . b ∈ X and all u ∈ U. a control sequence. t+N −1 the performance index J(¯. uk ). j ≥ 1. rst let u[t2 .t+N −1|t] .2) with N. Moreover. N ) x k=t subject to l(ˆk|t . the Finite Horizon Optimal Control Problem (FHOCP) consists in minimizing. ˆ x k ∈ [t. the nominal state dynamics 2.2. . i.1 Open-loop formulation In order to introduce the MPC algorithm formulated according to an openloop strategy. with t1 ≤ t2 ≤ k ≥ 0.k+j−1|k] . Given the ¯ Vf and the terminal Xf . ut3 |t1 ]. u) is such that l(0.5 (FHOCP) positive integer set Consider system (2. u[t. given xt = x. ¯ nd the optimal control sequence .4. uk|t ) + Vf (ˆt+N |t ) x x 1. there exists a positive constant Ll such that |l(a. u) ≥ αl (|x|) where αl is a K∞ -function. let xk+j|k be the ˆ with the nominal model f (xk . .4 The stage cost l(x. 0) = 0 and l(x. in the domain X × U . [ut2 |t1 ut2 +1|t1 . uk ). u) is Lipschitz in x.5). Nonlinear Model Predictive Control 23 2. xk+1 = f (ˆk . It is now possible to dene a prototype of a nonlinear MPC algorithm: at every time instants t. Assumption 2. with Then. (2. l(x.4). the following nite-horizon optimization problem can be stated. the terminal state constraints The stage cost denes the performance index to optimize and satises the following assumption. t + N − 1] xt+N |t ∈ Xf . the terminal penalty xt = x. with respect to u[t.e. u) − l(b. Moreover.4. u)| ≤ Ll |a − b| for all a. given predicted state at initial condition k + j obtained xk and input u[k. the stage cost l.t3 |t1 ] t3 . the constraints (2.

there exist a pair of suitable βVf such that αVf (|x|) ≤ Vf (x) ≤ βVf (|x|) 5.24 Chapter 2.19) uo [t. x. 3. 2. closed. given an auxiliary control law κf . Assumption 2. u = κM P C (x). Recall that. (2. κf (x))) − Vf (x) ≤ −l(x. Then. under suitable assumptions and by choosing accurately the terminal cost function Vf and the terminal constraint Xf . κf (xk )) is asymptotically function in Xf for the nominal system).19). αVf < βVf and x ∈ Xf K∞ -functions αVf and 4.t+N −1|t] is the solution of a constrained optimiza- tion.r. subject to constraints (2. Vf (f (x. according to the Receding Horizon (RH) strategy. since (2. The design parameters Vf and Xf are such that. stable in Xf (Vf is a .5 initial conditions.2) and (2. Vf is Lipschitz in for all x ∈ Xf LVf . κf (x)).t. dene κM P C (¯) x where uo (¯) t|t x and apply the control law uo (¯) t|t x is the rst column of uo [t. for all x ∈ Xf for all f (x. 6. κf (x)) ∈ Xf .t+N −1|t] . The solution of closed-loop system formed by (2.6 1. Xf κf (x) ∈ U .6 implies that the closed-loop system formed by the nominal system Lyapunov f (xk .2). Regional ISS for NMPC controllers uo [t. it is possible to guarantee the ISS property of the closed-loop system formed by (2. Although the FHOCP has been stated for nominal conditions. Xf with a Lipschitz constant Assumption 2. (2.t+N −1|t] by solving the FHOCP.3)-(2.19) could be discontinuous w.5). 0 ∈ Xf Xf ⊆ X .19) is continuous at x=0 ¯ and w=0 with respect to disturbances and Assumption 2.

19).19). Assumption 2.2.2) and (2.t+N −1|t] κf (ˆt+N |t+1 )] is an admissible.7 implies that closed-loop system (2. control sequence for the FHOCP at time ˜ xt ∈ X M P C (N ). The set Let where LJ N LVf Lf −1 + Ll N Lf −1 −1 Lf −1 . i. the optimal value of the performance index.20). Xf trary. A way to fulll it is shown in [Limon et al. IΩ.7 is not a-priori satised. . N ) (2.t+N −1|t] .6 (see e. w ∈ W. V (x) J(x. with the Many methods have been proposed in the literature to compute satisfying Assumption 2.20) is employed as an ISS-Lyapunov function for the closed-loop system formed by (2.4.17). In what follows.e. dened in (2.t+N |t+1] for all possible t + 1. Assumption 2. let X M P C (N ) denote the set of states for which a solution of the FHOCP problem exists. Assumption 2. is contained in W sup ).19). 2002a] by properly restricting the state constraints 2 and 3 in the formulation of the FHOCP.7 each Consider the closed-loop system (2. Remark 2. u[t+1. 2000]). uo [t. possible suboptimal.2) and (2. Nonlinear Model Predictive Control 25 In the following.8 • Ξ = X MP C • Ω = Xf • α1 = αl • α2 = βVf • α3 = αl • σ = LJ .2) and (2. For [uo x [t+1.g [Mayne et al. On the con- MPC based on the FHOCP dened above. X M P C (N ) is a RPIA set for the Note that Assumption 2.3 Vf . W is such that the set Θ (depending from with function V given by (2.

so that a less stringent approach explicitly accounting for the intrinsic feedback nature of any RH implementation has been proposed.t+N −1] .g. However. Gyurkovics 2002.3. the stage cost l − lw . the robust invariance of the feasible set X M P C (N ) in a standard open-loop MPC formulation can be achieved through a wise choice of the state constraints in the FHOCP.19) subject to constraints (2. . [Magni et al.2) with N. at any time instant the controller chooses the input u as a function of the current state x.2 Closed-loop formulation As underlined in Remark 2. see e.2) {l(xk . [κ0 (xt ) κ1 (xt+1 ) .1.4-2. RPIA set X ISS with 2.t+N −1] be stated. so as to guarantee that the eect of the disturbance w is compensated for any choice made by the nature.5) is M P C (N ). 2001b. Chen et al. Given ¯ penalty Vf and the the Xf .8 the closed-loop system formed by (2.2 Under Assumptions 2.26 Chapter 2.4. the state dynamics (2. Limon et al. the Finite Horizon Closed-Loop Game (FHCLG) problem consists in minimizing. κN −1 (xt+N −1 )] minimizing the cost in the worst disturbance case. .2) and (2. κ[t. at any time t the controller has to choose a vector of feedback control policies κ[t. Then. this solution can be extremely conservative and can provide a small RPIA set. In this framework. N ) x k=t subject to: 1.t+N −1] the cost function t+N −1 J(¯. Theorem 2.3)-(2. 1998.6 (FHCLG) positive integer terminal set spect to Consider system (2. w[t. the following optimal min-max problem can Denition 2.t+N −1] . 2003. with respect to κ[t. In the following. Hence. the terminal xt = x. uk ) − lw (wk )} + Vf (xt+N ) (2. 2006a]. 2. instead of optimizing with respect to a control sequence. Regional ISS for NMPC controllers The main stability result can now be stated.21) . Magni et al. it is shown that the ISS result of the previous section is also useful to derive the ISS property of min-max M P C .t+N −1] and maximizing with re- w[t.

0 ∈ Xf for all x ∈ Xf for all f (x. according to the RH paradigm. where Assumption 2.11 an auxiliary law 1.t+N −1] .22) κo (x) 0 is the rst element of κo [t. the feedback control law setting u = κM P C (x) is obtained by κM P C (x) = κo (x) 0 where (2. Nonlinear Model Predictive Control k ∈ [t.4. Assumption 2. (2. and all w∈W and 4. 3. given κf .t+N −1] be the solution of the FHCLG. 3.2. Xf ⊆ X . In order to derive the stability and performance properties associated to the solution of FHCLG.and all w ∈ W .19) could be κo [t.t+N −1] . κf (x))+w)−Vf (x) ≤ −l(x. αVf < βVf for all and x ∈ Xf . w[t. the constraints (2. .22) is continuous at x=0 ¯ and w=0 with respect to disturbances and Assumption 2. 2.9 initial conditions. that αw (|w|) ≤ lw (w) ≤ βw (|w|) . the following assumptions are introduced.t+N −1] is the solution of a constrained optimization. κf (x)) + w ∈ Xf . for all x ∈ Xf .3)-(2. x. 27 2.r.10 αw and βw are lw (w) is such K∞ -functions. Xf κf (x) ∈ U.5). Vf (f (x. t + N − 1] xt+N ∈ Xf . since (2. there exist a pair of suitable K∞ -functions αVf βVf such that αVf (|x|) ≤ Vf (x) ≤ βVf (|x|) x ∈ Xf 5.t.2). Recall that. κf (x))+lw (w). The design parameters Vf and Xf are such that. closed. discontinuous w. The solution of closed-loop system formed by (2. the terminal state constraints Letting o κo [t.

The main result can now be stated. dened in (2.5) is ISS with M P C (N ).12 the closed-loop system formed by (2. is contained in W sup ). κo [t.28 Chapter 2. In what follows.12 • Ξ = X MP C • Ω = Xf • α1 = αl • α2 = βVf • α3 = αl • σ = βw . Regional ISS for NMPC controllers Assumption 2. the optimal value of the performance index.11 implies that the closed-loop system formed by the system (2. The set Let W is such that the set Θ (depending from with function V given by (2. RPIA set X Note that.2) and (2.2) and in u = κf (xk ) is ISS in Xf (Vf is an ISS-Lyapunov function Xf ).17). Theorem 2. f (x. IΩ.3)-(2. 2.t+N −1] .t+N −1] .22) subject to constraints (2.3 Under Assumptions 2.e. u).23) is employed as an ISS-Lyapunov function for the closed-loop system formed by (2. i.9-2. lw and Vf . in this case.1. are not required to be Lipschitz. in order to prove the ISS.4. 2. l. Assumption 2.22). w[t.23).2) and (2. N ) (2. V (x) o J(x.

the control sequence ˜ u[t+1.11.4). This idea can either enlarge or restrict the set Ω1 ⊇ Ω but α2 ≥ α2 .12 for the To this regard. is an available feasible solution that guarantees ISS.2. In fact. assuming that an initial feasible solution of the FHOCP (or of the FHCLG) is available.t+N −1|t] (or κ[0. On be by the contrary. without assuming any regularity of the control strategies.5.30) and (2.6 tion V The usual way to derive the upper bound for the value funcM P C (N ) requires the assumption that the solutions uo in X [t. In this regard. is not trivial at all. it could be useful to nd an upper bound α2 ¯ of V in a region Ω1 ⊇ Ω. of the terminal penalty and of the terminal inequality constraint satisfying Assumption 2. Remarks Remark 2. it is possible to show that it is not necessary to obtain the global optimum solution of the FHOCP (or of the FHCLG) in order to guarantee the ISS of the closed-loop system. Theorem 2.N ] ). H∞ problem for the linearized 2. 1999] for standard NMPC. α2 = max ¯ where ¯ V . Indeed this solution is such that the value function satises (2. 2006a]. a solution for ane system will be discuss in Chapter 3.5 Remarks Remark 2. dene FHOCP and FHCLG respectively. where it will shown how to compute a nonlinear auxiliary control law based on the solution of a suitable system under control.t+N |t+1] [˜ [t+1.t+N |t+1] (or than κ[1. in order to enlarge the set using the monotonicity properties (2. However. This can derived in Xf .t+N −1] of the FHOCP and the κo [t.8 and 2. ¯ . 1 α2 α2 (r) Br = {x ∈ Rn : |x| ≤ r} ⊆ Ω.t+N −1|t] κf (ˆt+N |t+1 )] (or the vector of feedback conu x ˜ trol policies κ[1. W that satises Assumptions 2.N ] ˜ [˜ [0. ¯ V = maxx∈Ω1 (V (x)) and r W since is such that as suggested in [Limon et al. where u[t+1.N −1] κf ]).5 Following the result in [Scokaert et al.t+N −1] of the FHCLG are Lipschitz in X M P C (N ). The only requirement on the possible suboptimal solution is to be not worst than ˜ u[t+1.4 29 The computation of the auxiliary control law.1 gives the possibility to nd the upper bound for ISS -Lyapunov function only in a subset of the RPIA set.34) respectively.N −1] ) is the κ ˜ possible sub-optimal solution obtained at the previous step. ˜ Remark 2.

i. Moreover an estimation of the region where the state of the system converges asymptotically is given. non necessarily continuous. 2. x.7 Appendix Proof of Theorem 2.14). This result has been used to study the robustness characteristics of Model Predictive Control algorithms derived according to open-loop and closed-loop min-max formulations. assume that there exists a nite time t x(t.7). w) ∈ Θw .e. this implies ¯ ¯ ρ ◦ α4 (V (x(t. Using (2.30 Remark 2. such that for To this aim. Moreover the robust output admissible set guaranteed with a longer optimization horizon includes the one obtained with a shorter M P C (N + 1) ⊇ X M P C (N ) ⊇ X .7 Chapter 2. w)) ≤ b(||w||).1: the proof will be carried out in three steps. 2003] for the closed-loop min-max MPC formulations. it is going to be shown that Θw is a RPI set system (2. Regional ISS for NMPC controllers Following the results reported in [Magni et al. Then V (x(t.15) can be rewritten as ¯ ∆V (x) ≤ −α4 (V (x)) + σ(|w|). Lyapunov function. regional Input-to-State Stability for discrete-time nonlinear constrained systems has been established. Without loss of generality.6 Conclusions Using a suitable. No continuity assumptions on the optimal Receding Horizon control law have been required. w)) ≤ σ(||w||). x. 2001a] for the open-loop and in [Magni et al. (2. It is believed that the elements provided here can be used to improve the analysis of existing MPC algorithms as well as to develop new synthesis methods with enhanced robustness properties. X f 2. assume that (id − α4 ) is . Step 1: rst. it is easy to show that the robust output admissible sets guaranteed by the NMPC control law include the terminal region Xf used in the optimization problem. The next chapters will be based on the fundamental results here presented. ∀w ∈ W where −1 α4 = α3 ◦ α2 . x. horizon. ∀x ∈ Ω.

since α4 = α3 ◦ −1 α2 . V (x(t + 1. Considering that one has ρ ◦ α4 ◦ b(||w||) = σ(||w||) and (id − ρ) is a K∞ -function. By the fact that −1 ¯ ¯ α2 (V (x(k. w)) ≤ (id − α4 )(V (x(t.2. x. starting from asymptotically to Θw . w)) > ρ ◦ α3 (|x(k. w))) > σ(||w||). x. ∀x ∈ Ω \ Θw ¯ ¯ which implies −α3 (|x(k. w)|) > σ(||w||). x. w))) + σ(|w|) ¯ ¯ ≤ (id − α4 ) ◦ b(||w||) + σ(||w||) = −(id − ρ) ◦ α4 ◦ b(||w||) + b(||w||) −ρ ◦ α4 ◦ b(||w||) + σ(||w||). x. Appendix K∞ -function. w)) ≤ b(||w||) for all j ∈ Z≥0 . w)|) + σ(||w||) < 0. x. ¯ Moreover. x. ∀x ∈ Ω \ Θw . x. x. w)|. the state tends x ∈ Ω \ Θw then Step 2: now it is shown that. x. by denition of (2. x. there of Denition 2. ¯ By induction one can show that that is (2. x. x(k. ∀s > 0 then α3 (x(k. one has ρ ◦ α3 (|x(k. w)) ≤ −(id − ρ) ◦ α4 ◦ b(||w||) + b(||w||) ≤ b(||w||). α2 α3 < α2 . w)|) > σ(||w||). if −1 ¯ ρ ◦ α3 ◦ α2 (V (x(k. w))) ≤ |x(k.7).24) c > 0 ¯ such that for all Θw (see point 4 x ∈ Ξ \ Ω. Firstly. x. w) ∈ Θw ¯ for all V (x(t + j. there exists exists x ∈ Ω \ Θw ˜ such that . Hence Θw is a RPI set for system Ξ \ Θw . 31 a otherwise take a bigger so that Then V (x(t + 1. ¯ k ≥ t. ¯ Considering that (id − ρ) is a K∞ -function id(s) > ρ(s).7.4).

for x(T1 . ¯ ¯ / ρ ◦ α4 (V (x(T1 . x. w)) < −c2 . x. x.7) is regional ISS in Ξ. w)|Θw = 0. w)) ≤ ε + b(||w||). ∀x ∈ Ξ \ Ω. that means. x. there exists a c > 0 such that Θw = c. the state will reach the region Ω in a nite time. w)| ≤ α1 (ε + b(||w||)). x. x. there is |x(T1 . ¯ . starting from Θw in a nite time and in Θw asymptotically. Regional ISS for NMPC controllers α3 (|˜|) ≤ α3 (|x|) − c. x ¯ Then from (2. Hence.26) and (2. ¯ sup Note that if ||w|| = 0. x. Step 3: nally it is shown that system (2. ¯ x Then ∆V (x(k. Otherwise. there exists T1 such that x(T1 . x.25) ∀ε > 0. w)|) ¯ where the last step is obtained using (2. Hence limk→∞ |x(k. x. w) ∈ Θ. ¯ Ξ. x. w)) ≤ −α4 (V (x(T1 . w))) > σ(||w||) and ¯ Therefore. x.13) one has −1 |x(T2 . the ¯ / ¯ state arrives in Θw in a nite time. w) ∈ Ω. Then. x. the state will arrive close to Ξ. ∆V (x(T1 . If in particular x(T1 . w)) ≤ −α3 (|x(k. w) ∈ Θw . x.13). ¯ Therefore.26) V (x(T2 . w)|) + σ(||w||) < −¯. Then. x. then the previous result states ¯ that the region Θw is achieved in a nite time. w) ∈ Θw . x. x. w))) + σ(||w||) ¯ ¯ = −(id − ρ) ◦ α4 (V (x(T1 .32 Chapter 2. w)| > c. w))) ¯ −ρ ◦ α4 (V (x(T1 . there exists a c2 > 0 ¯ / ¯ ˜ such that. x. starting from ∆V (x(T1 . w))) + σ(||w||) ¯ ≤ −(id − ρ) ◦ α4 (V (x(T1 . w)|Θw = 0.24) it follows that −α3 (|x|) + c ≤ −α3 (|˜|) < −σ(||w||). for ˜ ˜ x(T1 . ∀x ∈ Ξ \ Ω ¯ ¯ c so that. Since Θw is a RPI set. such that (2. ∃T2 (ε ) ≥ T1 (2. By (2. it is true that limk→∞ |x(k. if x(T1 . x. considering that −1 |x| ≤ α1 (V (x)). x. w))) ¯ ≤ −(id − ρ) ◦ α4 ◦ α1 (|x(T1 . w) ∈ Θw . x.

x. it follows that ¯ J(x. N + 1) ≤ J(x. 2006a]. u[t. Then (2. N ) + Vf (f (ˆt+N |t .2. ¯ Now. κf (ˆt+N |t )).2.t+N −1|t] κf (ˆt+N |t )] is an admissible. x x x Using point 5 of Assumption 2. N + 1) = V (x. it follows that −1 −1 |x(T2 .t+N |t] [uo x [t. ∀x ∈ X M P C (N ). V (x) dened in (2.6. Moreover. ∀x ∈ Xf .20) M P C (N ).28) V (x. letting in −1 ε = α1 (2ε ) and −1 γ(||w||) = α1 (2b(||w||)) the UAG property Ξ is proven. κf (ˆt+N |t ))) x x −Vf (ˆt+N |t ) + l(ˆt+N |t . the lower bound is easily obtained using Assumption 2. ∀x ∈ X M P C (N ) with (2. in view of Assumption 2. N + 1). N ) ≥ l(x. it will be shown that the function X M P C (N ).t+N |t] . Therefore.1. 33 Noting that. N + 1) ≤ V (x.2: ISS-Lyapunov function in by Theorem 2. N + 1) ≤ V (x.27) ¯ u[t. if system admits an then it is ISS in Proof of Theorem 2. possible suboptimal.t+N |t] is a suboptimal sequence ¯ V (x. θ1 (s1 + s2 ) ≤ θ1 (2s1 ) + θ1 (2s2 ). u[t.4 and considering that V (x. Since ¯ u[t. is an ISS-Lyapunov function in X First. control sequence for the FHOCP with horizon N +1 at time t with cost ¯ J(x. the system is ISS in Ξ with respect to w. N ). w)| ≤ α1 (2ε ) + α1 (2b(||w||)). using point 4 of Assumption . In the following. u[t.2 and 2. 0) = Vf (x).6 (2. X M P C (N ).3 and Lemma 2. In view of Assumptions 2.t+N |t] .29) V (x. given a see [Limon et al. κM P C (x)) ≥ αl (|x|).7. N ).t+N |t] . Appendix K∞ -function θ1 .

Dene ∆J as ∆J ˜ J(xt+1 .4. u[t.31) .t+N |t+1] to the nominal state xt+1 at time t + 1.6. N − 1) ≤ V (x. x x x x From the denition of ˜ ˜ u. uk|t+1 ) − l(ˆk|t . Denote xk|t+1 ˜ model. N ) ≤ ∆J ≤ −αl (|xt |) + LJ |wt | (2. and hence.29) with sequence u.28) and (2. uo )| ≤ Ll Lk−t−1 |wt |. uo ) + t|t k=t+1 {l(˜k|t+1 . N ) − J(x(t). κf (ˆt+N |t ))). by Assumptions 2. uo ) + t|t k=t+1 Ll Lk−t−1 |wt | f +Vf (˜t+N |t+1 ) − Vf (ˆt+N |t ) x x Moreover.t+N |t+1] .1 and 2.7. N + 1) ¯ t+N −1 = −l(xt . ˜ the bound on the decrease of V is V (xt+1 .30) ˜ u[t+1.1 and 2. by using point 6 of Assumptions 2. at both x ˜ and x ˆ t+N −1 ∆J ≤ −l(xt . t + N − 1]. ∀x ∈ Xf . u[t+1.34 Chapter 2. x ˜ x k|t f Then. N ) − V (xt . 0) = Vf (x) ≤ βVf (|x|) . there is ∆J ≤ −l(xt . the state obtained at time starting from the real ˜ k applying u[t+1. uo ) + LJ |wt | ≤ −αl (|xt |) + LJ |wt | t|t where LJ u ¯ Lv LN −1 + Ll f LN −1 −1 f Lf −1 . Finally. uk|t+1 ) − l(ˆk|t . κf (ˆt+N |t )) x x x x +Vf (f (˜t+N |t+1 . uk|t )} x ˜ x ¯ +l(˜t+N |t+1 . let (2. N ) ≤ V (x.t+N |t+1] be as dened in Assumption 2. Regional ISS for NMPC controllers 2. x x f Substituting these expressions in ∆J .t+N |t] . and by Assumption 2. Now. κf (˜t+N |t+1 )) − l(ˆt+N |t .4 |l(˜k|t+1 .6 |Vf (˜t+N |t+1 ) − Vf (ˆt+N |t )| ≤ Lv LN −1 |wt |. ¯ for k ∈ [t + 1. by applying also point 5 of Assumption 2. uk|t+1 = uk|t . the upper bound is obtained V (x. κf (˜t+N |t+1 ))) − Vf (f (ˆt+N |t . by using (2.29) with sequence obtained and (2.6.

N ) ≥ κ[t. κ[t.23) is an ISS-Lyapunov function for the closed-loop system (2.2). N + 1) = ˜ k=t {l(xk .t+N ] . (2.5.32) x ∈ X M P C (N ).1. Appendix x ∈ X M P C (N ) w ∈ W. w[t.t+N −1|t] κf (xt+N )] as admissible policy o from the optimal sequence κ [t.11 by taking vector at time κ[t+1. then it is ISS in X M P C (N ).7.t+N −1] at time t. κ[t.t+N −1|t] . w[t.t+N −1] κf (xt+N )] as admissible t with horizon N + 1. uo [t. w[t. the lower bound is easily obtained o V (x.2.2). N ) M P C (N ) and system in X hence. consider the following policy vector κ[t. uk ) − lw (wk )} +Vf (xt+N ) − Vf (xt+N ) + Vf (xt+N +1 ) +l(xt+N .11 t+N −1 J(x. uk ) − lw (wk )} + Vf (xt+N ) .t+N −1] . (2. (2.t+N ] ˜ at time J(x. 0.31) the optimal cost is an ISS-Lyapunov function for the closed-loop J(x. 35 for all and all Therefore. ut+N ) − lw (wt+N ). First. κ0 (x)) ≥ αl (|x|) for all (2. κ[t.30).3: ISS-Lyapunov function in X M P C (N ).t+N ] .t+N −1] .t+N ] .t+N −1] min J(¯. the closed-loop system (2.19) is ISS with RPIA set X M P C (N ). the robust invariance of 2. In view of Assumption 2. κo [t.t+N ] . N + 1) ≤ ˜ k=t {l(xk . by (2. if system admits an Proof of Theorem 2. Then. Then t+N −1 policy vector for the FHCLG In order to derive the upper bound. (2. considering also Assumption 2. [κo [t. In the following it will be shown that the function V (x) dened in (2.22) in X M P C (N ). N ) x ≥ l(x.27).t+N −1] . N ) = J(x.t+N ] ¯ t + 1 starting X M P C (N ) is easily derived from Assumption [κo [t+1. by Theorem 2.

[w w[t+1. κM P C (x)) + w. κM P C (x)) + w. N − 1) ≤ . . N ) − V (x.33) and all = V (x.t+N −1|t+1] ]. N ) ≤ −l(x. considering also Assumption 2.t+N ] . ≤ V (x. the bound on the decrease of V is obtained V (f (x. N ) which holds for all x ∈ X M P C (N ) w ∈ MW .t+N −1|t+1] .34).35). N + 1) ˜ t+N −1 w∈MW max {l(xk . the optimal and all w ∈ W.2). N − 1) − V (x.4 and 2. by applying Assumptions 2.t+N −1|t] . N + 1) ≤ ≤ w∈MW max J(x. uk ) − lw (wk )} + Vf (xt+N ) k=t (2.32). κo [t. 0) = Vf (x) ≤ βVf (|x|) for all (2. using Assumption 2. κM P C (x)) + w. N ) ≤ V (f (x. the closed-loop system (2.22) is ISS with RPIA set X M P C (N ).t+N ] . w[t. N − 1) o − J(x. Regional ISS for NMPC controllers which implies V (x. w[t+1. N ) ≤ V (x. Therefore. N ) o ≤ J(f (x. N ) ≤ −αl (|x|) + βw (|w|) for all (2. N ) − V (x. (2. (2.9.11. κM P C (x)) + w. (2. function for the cost closed-loop system in V (x) M P C (N ) X ISS-Lyapunov and hence. From the monotonicity property (2.35) x ∈ X M P C (N ). .t+N −1|t] . the upper bound is obtained V (x. by (2.10.34) x ∈ Xf . . is an Therefore. κo [t+1. κM P C (x)) + lw (w) so that. κ[t.36 Chapter 2.33) and by sub-optimality there is V (f (x.

Chapter 3

Min-Max NMPC: an overview on stability

Contents

3.1 3.2 3.3 3.4 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 39 42 47 Problem statement Regional Input-to-State practical Stability

Min-max model predictive control . . . . . . . . . .

3.4.1

3.5 3.6 3.7 3.8

Stability . . . . . . . . . . . . . . . . . . . . . . . . .

49

55 57 62 62

The auxiliary control law . . . . . . . . . . . . . . . . Example . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusions . . . . . . . . . . . . . . . . . . . . . . . . Appendix . . . . . . . . . . . . . . . . . . . . . . . . .

3.1 Introduction

Min-Max model predictive control (MPC) is one of the few techniques suitable for robust stabilization of uncertain nonlinear systems sub-

ject to constraints.

Stability issues as well as robustness have been

recently studied and some novel contributions on this topic have appeared in the literature [Scokaert & Mayne 1998, Gyurkovics 2002,

Kerrigan & Mayne 2002, Fontes & Magni 2003, Lazar et al. 2008,

Magni et al. 2003,

Gyurkovics & Takacs 2003, Lazar 2006, Löfberg 2003,

Magni & Scattolini 2005, Magni et al. 2006a,

Limon et al. 2006a,

Kerrigan & Maciejowski 2004].

In this chapter, a general framework for

synthesizing min-max MPC schemes with an a priori robust stability

38

Chapter 3. Min-Max NMPC: an overview on stability

guarantee is distilled starting from an extensive literature. To this aim, a general prediction model that covers a wide class of uncertainty modeling that includes bounded disturbances as well as state (and input) dependent disturbances (uncertainties) is introduced. This requires that the regional Input-to-State Stability (ISS) and Input-to-State practical Stability (ISpS) results are extended in order to cover both state dependent and state independent uncertainties. Moreover, Lyapunov-type sucient conditions for the regional ISS and ISpS are presented for the considered class of systems; this constitutes the base of the stability analysis of the min-max MPC for the generalized prediction model. Using the ISS tool, it is proven that if the auxiliary control law ensures ISS of the system, with respect to the state independent part of the disturbance (i.e. the state dependent part of the disturbance remains in a

certain stability margin) then the min-max MPC ensures that the closed loop system is ISpS maintaining the same stability margin. The practical nature of the stability is a consequence of the worst-case approach of the control action and causes the system to be ultimately bounded even if the real disturbances vanish. In order to avoid this problem, two dierent possible solutions are considered: the rst one is based on an

H∞

like cost of the performance

index (see e.g. [Magni et al. 2001b, Magni et al. 2003, Magni et al. 2006a, Gyurkovics & Takacs 2003]), while the second one is based on a dual-mode strategy [Michalska & Mayne 1993, Chisci et al. 1996, Scokaert et al. 1999, Scokaert & Mayne 1998, Kerrigan & Mayne 2002, Lazar et al. 2008].

These solutions are extended to a general prediction model and it is shown that under fairly mild assumptions both controllers guarantee Input-to-State Stability. Moreover, a nonlinear auxiliary control law, based on the one presented in [Magni et al. 2003], is proposed for the case of nonlinear systems ane in control (which are very usual). It is shown that a nonlinear auxiliary

control law and the terminal penalty can be derived from the solution of the discrete-time

H∞

algebraic Riccati equation for the linearized system.

3.2. Problem statement

39

3.2 Problem statement

Assume that the plant to be controlled is described by discrete-time nonlinear model

**xk+1 = f (xk , uk , d1k , d2k ), k ≥ 0
**

where

(3.1)

f : IRn × IRm × IRp × IRq → IRn is a nonlinear, possibly discontinuous n m function, xk ∈ IR is the system state, uk ∈ IR is the current control vector and d1k and d2k are disturbances which model the uncertainties present in

the model. This partition on the disturbance signals stems from its nature:

d1k ∈ IRp

models a class of uncertainty which depends on the state and the

control input while

d2k ∈ IRq

models a class of uncertainty that does not

depend neither on the state nor on the input signal. The transient of the system (3.1) with initial state is denoted by

x0 = x ¯

and disturbance sequences

d1

and

d2

x(k, x, d1 , d2 ). ¯

Most of the models of nonlinear systems considers the uncertainty as bounded disturbances, that is, the only knowledge of the model mismatches is a bounded set where the error lies in. However, this representation may lead to conservative results when, as usually occurs, there exists a relationship between the model mismatch bounds and the state and input of the plant. In this case, this conservativeness would be reduced if this information were considered in the model of the plant by means of the proposed partition of the disturbance model. In the following assumption, the considered structure of such models is formally presented.

Assumption 3.1

1. The system has an equilibrium point at the origin, that is

f (0, 0, 0, 0) =

0.

2. The uncertainty

d1

is such that

d1 = d1η η(|(x, u)|)

where

(3.2)

η

is a known

K-function

and

d1η ∈ IRp

**is modeled as conned
**

(3.3)

in a compact set

D1η ⊂ IRp

not necessarily including the origin with

inf D1η

and

sup D1η

known.

40

**Chapter 3. Min-Max NMPC: an overview on stability
**

d2

3. The uncertainty

is such that

d2 ∈ D2

where

(3.4)

D2 ⊂ IRq

is a compact set containing the origin with

sup D2

known. 4. The state and control of the plant must fulll the following constraints on the state and the input:

x∈X u∈U

where

(3.5) (3.6)

X

and

U

are compact sets, both of them containing the origin

as an interior point. 5. The state of the plant

xk

can be measured at each sample time.

The control objective consists in designing a control law

u = κ(x)

such

that it steers the system to (a neighborhood of ) the origin fullling the constraints on the input and the state along the system evolution for any possible disturbance and yielding an optimal closed-loop performance according to certain performance index. This control problem is well studied in the literature and there exist a number of control techniques that could be used. However, among the

existing solutions, one of the most successfully used control technique is the model predictive control in its min-max approach. This is due to its optimal formulation, its capability to ensure the robust constraint satisfaction and its stabilizing design [Mayne et al. 2000]. In the min-max MPC controllers, the control law is based on the solution of a nite horizon game, where (the controller), and

u

is the input of the minimizing player

d1η , d2

are the maximizing player (the `nature'). More

precisely, the controller chooses the input state

uk

as a function of the current

xk

so as to ensure constraint satisfaction along the predicted trajectory

of the plant for any possible uncertainty, minimizing at the same time the worst case performance index of the predicted evolution of the system. In the open-loop min-max MPC strategy, the performance index is minimized with respect to a sequence of control action, a sequence which belongs

3.2. Problem statement

41

to a nite-dimensional space. It is well known that it is not convenient to consider open-loop control strategies since open-loop control would not account for changes in the state due to unpredictable inputs played by `nature' (see also [Scokaert & Mayne 1998]). If a control law is considered as decision variable in the optimization problem (instead of a control action), the solution results to be less conservative since the predicted controlled system reacts to the eect of the disturbance. The predictive controllers derived from this approach are called

**closed-loop min-max MPC controllers and can provide larger domain of
**

attraction and a better closed-loop performance index. However, the optimization problem may be dicult (or even impossible) to be solved even for linear prediction models. A practical solution, located between the open-loop and the closed-loop approach, is the so-called semi-feedback formulation of the problem. In this case, control policies are considered as decision variables, but forcing a given structure of the control law. Thus, the decision variable of each control law is its set of dening parameters, yielding to an optimization problem similar to the open-loop case one. In this chapter, the considered control law is derived from a closed-loop min-max MPC formulation

κ(x) = κM P C (x).

Although, from a practical

point of view, the control law is dicult to calculate, from a theoretical point of view makes sense since the closed-loop approach includes the open-loop and semi-feedback controllers, and these can be considered as particular cases. Thus, the stability results derived in the following for closed-loop It is worth re-

min-max MPC will be valid for the rest of formulations.

marking that this control law might be a discontinuous function of the state. The resulting closed-loop system is given by

**xk+1 = f (xk , κM P C (xk ), d1k , d2k ), k ≥ 0
**

where the disturbance

d1k = d1ηk η(|(xk , κM P C (xk ))|) with d1η ∈ D1η , and the disturbance d2k is such that d2k ∈ D2 . The control law should ensure that if the initial state is in a certain RPIA set, i.e. xt ∈ X M P C , then the resulting evolution of the system fullls the constraints, M P C (x ) ∈ U for all k ≥ t, for any possible evolution that is xk ∈ X and κ k d1k

is such that of the disturbance signals.

2. The transient of the system (3.3 in Chapter 2. Dierently from section 2. 0) ∈ A. does not converge to the origin but just to a neighborhood of it.7) with initial state x0 = x ¯ and uncertainties d1 and d2 is denoted by x(k. 0. that means F (x. As it will be explained. This system is ¯ supposed to fulll the following assumption. d2 ). 0. Consider a nonlinear discrete-time system described by xk+1 = F (xk . Min-Max NMPC: an overview on stability In the following section it is presented a suitable framework for the analysis of stability of such class of closed loop systems: the regional ISpS. a positively invariant set for the associated undisturbed system xk+1 = F (xk .42 Chapter 3. containing the origin. x. d2k ).7) F : IRn × IRp × IRq → IRn is a nonlinear possibly discontinuous n p function. d1k ∈ IR is the component of the uncertainty q depending from the state and d2k ∈ R is the other component of the uncertainty. The uncertainty d1 is such that d1 ∈ D1 (x) ⊆ IRp where. that is. ∀x ∈ A. this tool copes also with systems that.3 Regional Input-to-State practical Stability In this section the ISpS framework for discrete-time autonomous nonlinear systems is presented and Lyapunov-like sucient conditions are provided. 0). d1k . This will be employed in this chapter to study the behavior of perturbed nonlinear systems in closed-loop with min-max MPC controllers.7). D1 (x) is closed and contains the origin. xk ∈ IR is the state. Moreover . k ≥ 0 where (3. this is for example the case of closed-loop systems controlled with a standard min-max MPC law. The compact set A ⊂ IRn . d1 . Assumption 3. 3.2 1. is a zero-invariant set for the system (3. for each x. even in absence of disturbances.

x. i. In many applications it is of interest to study the stability with respect to an invariant set a single point. version of Input-to-State practical Stability (ISpS) regional [Sontag & Wang 1996. c = 0. and the origin is an equilibrium point for undisturbed system. Denition 3. Gao & Lin 2000]. a K-function γ2 and a constant c ≥ 0 such that Given a compact set |x(k. presented [Sontag & Wang 1995a. the system (3. that .7) and if there exist a KL-function β . is given in the Appendix of the Thesis. 0) = 0.4 of [Sontag & Wang 1995b].e.1 (ISpS in Ξ) Ξ ⊂ IRn . A. where A in general does not consist of The set version of the ISS property (also known A in as compact-ISS) was originally proposed in [Sontag & Lin 1992]. Regional Input-to-State practical Stability K-function η and a signal d1η ∈ IRp .8) x∈Ξ ¯ and k ≥ 0.8) is satised with Denition 2. F (0. the system (3.3 of [Sontag & Wang 1996] and in Denition 2. Jiang & Wang 2001] is dened in the following.7) is said to be ISS (Input-to-State Stable) in Ξ with respect to d2 (see Note that.3. it is shown. 3. d1 .3. including the origin as an interior point. whenever (3. k) + γ2 (||d2 ||) + c ¯ x for all (3. limited in a ⊂ IRp (not necessarily including the origin as an d1 = d1η η(|x|) for all 43 there exist a set compact interior D1η point) such that x ∈ Ξ. The uncertainty d2 is such that d2 ∈ D2 ⊂ IRq where A D2 is a compact set containing the origin. 0. d2 )| ≤ β(|¯|. where Ξ ⊂ IRn is a compact set containing the origin as an interior point. regional version of the global ISS with respect to A.3). In Proposition VI.7) with d1 ∈ MD1 and d2 ∈ MD2 is said to be ISpS (Input-to-State practical Stable) in Ξ with respect to d2 if Ξ is a RPI set for (3. for the continuous-time case.

In fact ISS in Ξ with respect Ξ and ISpS in Ξ with A ≡ Ξ are always satised.1 to The assumption that Ξ is a RPI set could render the de- Ξ trivials. as discussed in Chapter 2. that extends the ISS results of [Magni et al. if a RPI compact set Ξ is considered.2 is satised. one can see that the equivalence between ISS in Ξ and the conjunction of UAG and LS (and consequently the equiv- alence between ISpS in Ξ and the conjunction of UpAG in Ξ and LpS) also applies to discontinuous systems (both continuous and discrete-time).7 in [Sontag & Wang 1995a] carefully. However. using the results of [Limon et al. A sucient condition.7). Summarizing. The following properties are equivalent Ξ with respect to Ξ A c) UpAG in d) UAG in Ξ and LpS Ξ with respect to A and LS with respect to A. on the base of the results in [Sontag & Wang 1995a. . Theorem 3. concept of ISpS is equivalent to concept of ISS with respect to case.8) is Asup . a priori non-smooth) with respect to d2 . if in the ISpS property. Sontag & Wang 1996. As for the ISpS property. the constant In this c of equation (3. it is possible to give the following equivalence result. Sup- pose that Assumption 3. Moreover.1. 2006a]. Min-Max NMPC: an overview on stability A.44 Chapter 3. In order to clarify the relation between the sets introduced in the following denition see Figure 3. Gao & Lin 2000]. nitions of ISpS and UpAG in Regional ISpS will be now associated to the existence of a suitable Lyapunov function (in general. c is shown to be smaller than Ξsup . Sontag & Wang 1995b.1 a) ISpS in b) ISS in Let Ξ ⊂ IRn be a compact set. Remark 3. then the ISpS in Ξ (or the ISS in Ξ with respect to A) give more information than the solely robust positively invariance of Ξ. Consider system (3. is introduced. by examinating the proof of Lemma 2. the concepts of LpS and UpAG are equivalent to the concepts of LS and UAG with respect to A (see the Appendix of the Thesis for the denitions of the properties). 2006a]. The same relation holds for discrete-time case. and in the ISS the set A is shown to be smaller than Ξ.

3. a pair of suitable such that Ω ⊆ Ξ (including the origin as an interior K∞ -functions α1 .10) V (x) ≥ α1 (|x|). c1 = c2 = 0. whenever Denition 3. α2 and a constant c1 ≥ 0 (3. given a disturbance sequence Θd2 ⊆ IΩ d2 ∈ MD2 . the Note that. one has to verify that (3.2 1. d1 . 45 V : IRn → IR≥ 0 (ISpS-Lyapunov function in Ξ) A function is called an ISpS-Lyapunov function in Ξ for system (3. V is an ISS-Lyapunov function in Ξ for system (3.7). such that. α2 α2 (s) + s. there exist a suitable K∞ -function α3 . there exist suitable is a set K∞ -functions ζ and K∞ -function) and a suitable constant ρ such that (id − ρ) cθ > 0. there exist a compact set point). with α4 min(α3 (s/2).4). as discussed in Chapter 2. d2 ∈ D2 ρ (with 4. |x|δΩ > cθ } exists a nonempty compact (including the origin as an interior point) dened as follows: Θd2 where {x : V (x) ≤ b(λ2 (||d2 ||) + c3 )} (3. a K-function λ2 and a constant c2 ≥ 0 such that ∆V (x) V (F (x. c3 α3 ◦ α−1 . if Ξ is a compact RPI set including the origin as an interior point 2. there {x : x ∈ Ω.9) (3.2 is satised with function 2. d2 )) − V (x) ≤ −α3 (|x|) + λ2 (|d2 |) + c2 and all (3. α3 (s) 2 c2 + ζ(c1 ).3.2 Θd2 ⊆ IΩ Note that. ∀x ∈ Ξ V (x) ≤ α2 (|x|) + c1 .11) for all x ∈ Ξ. Regional Input-to-State practical Stability Denition 3.12) −1 b α4 ◦ ρ−1 . ∀x ∈ Ω 3.7) (see Denition Remark 3. all d1 ∈ D1 (x). in order to verify that for all d2 ∈ MD2 . ζ(s/2)).13) Θ sup {x : V (x) ≤ b(λ2 (D2 ) + c3 )} ⊆ IΩ .

x. and all (3. If system (3. d1 .3 Note that for a generic disturbance d1 . . all d2 ∈ D2 . for limk→∞ |x(k. Min-Max NMPC: an overview on stability Figure 3.7).2 holds. condition 3 of Deni- tion 3. d1 . c2 ≥ 0 such that a pair of K-functions λ1 and ∆V (x) V (F (x. Theorem 3. a sucient condition for regional ISpS of system (3.1: Example of sets satisfying Denition 3.2 Now. Remark 3. d2 )) − V (x) ≤ −α(|x|) + λ1 (|d1 |) + λ2 (|d2 |) + c2 d1 ∈ D1 (x). then it is d2 and. 2006a] using the results of [Limon et al.46 Chapter 3. 2006a]. ¯ Ξ with respect to all disturbance sequences ISpS d2 ∈ MD2 .7) ad- mits an ISpS-Lyapunov function in Ξ with respect to d2 . that extends the ISS results of [Magni et al.2 in Suppose that Assumption 3.14) for all x ∈ Ξ. d2 )|Θd2 = 0.2 should be 3. there exist a suitable λ2 and a constant K∞ -function α3 . can be stated.

Given the ¯ penalty Vf and the ter- .7) must have a stability margin: under Assumption 3. This optimization problem can be posed as the following Finite Horizon Closed-Loop Game (FHCLG). in order to verify that point 4 of Denition 3.2.1) with N. the function 3.15) In fact. Firstly. This means that system (3. 47 However. the control law derived by means of a closed-loop min-max MPC considers a vector of feedback control policies κ[t. In order to clarify the 3. Min-max model predictive control d1 x.2. since is a function of the term in (3.2. the stability of dierent approaches of this control technique is studied.2 gives an estimation of the region Remark 3. the terminal xt = x. it is required that −α(|x|) + λ1 (|d1 |) = −α(|x|) + λ1 (|d1η |η(|x|)) sup ≤ −α(|x|) + λ1 (D1η η(x)) with −α3 (|x|) α3 ∈ K∞ -function. ¯ Θ ⊇ Θ.2.9) and Remark 3.t+N −1] [κ0 (xt ). Then.2.4 Min-max model predictive control This section presents new results that allows the presentation of previous results in a unied framework. κN −1 (xt+N −1 )] in the minimization of the cost in the worst disturbance case. in view of Assumption 3.4. the formulation of the closed-loop min-max control law is presented. Theorem 3. Hence. one has to verify that V is not known in explicit form. deriving sucient conditions and generalizing existing results. κ1 (xt+1 ). Denition 3. see Figure 3. in view of (3. the stage cost l. it is necessary that the K∞ -function α compensates the eect of the disturbance d1 .2 is satised. As it was claimed in Section 3. considering also Remark ¯ Θ sup −1 {x : |x| ≤ α1 ◦ b(D2 )} ⊆ IΩ (3. relation between the sets. · · ·. In some cases. In order to satisfy the condition K∞ -function.3.4 Θd2 where the state of the system converges asymptotically.14) λ1 (d1 (x)) that α3 is is incorporated in a −α3 (|x|). as for example in the MPC.3 (FHCLG) positive integer Consider system (3.2.

t+N−1] .t+N−1] be the solution of the FH- CLG. d1[t. κ[t.6). the state dynamics (3.2)-(3. the FHCLG problem consists in minimizing.t+N −1] .t+N −1] and maximizing with respect to d1[t.t+N−1] the cost function J(¯.17) . d1[t. according to the Receding Horizon (RH) paradigm. Min-Max NMPC: an overview on stability Figure 3.48 Chapter 3. the constraints (3. d2[t.t+N −1] .t+N−1] . the terminal constraint In the following. d2k ) + Vf (xt+N ) k=t subject to 1.2: Relation between sets minal set Xf .t+N−1] . t + N − 1] xt+N ∈ Xf .t+N−1] and d2[t. uk . let (3. d2[t. the feedback control law u = κM P C (x) is obtained by setting κM P C (x) = κo (x) 0 (3.1) 2.16) k ∈ [t. d1k . 3. Letting o o κo [t. denote the set of states for which a X M P C (N ) solution of the FHCLG problem exists. N ) x t+N −1 l(xk . with respect to κ[t.

3. The design parameters Vf . when persistent disturbances are present. 2000]. the stage cost function.1 Stability In this section. d2 ) ∈ Xf . d2 ∈ D2 4. and As it is standard in MPC [Mayne et al. Min-max model predictive control κo (x) 0 κo [t. it will be shown that.t+N −1] .3 The stage cost l(·) is such that l(x. all where Lκf > 0 and all for all x ∈ Xf . 0. two dierent solutions for overcoming this problem and guaranteeing ISS of the min-max MPC closed-loop system are given: the rst one is derived using a particular design of the stage cost of the performance index. The stage cost denes the performance index to optimize and must satisfy the following assumption. d2 ) ≥ αl (|x|) − αd (|d2 |) where αl and l(0. 3. αVf < βVf for all and x ∈ Xf . κf (x). In order to derive the main stability and performance properties associated to the solution of FHCLG. the terminal cost function and the terminal region. Secondly. the standard min-max approach can only guarantee ISpS. d1 .4. the following assumption is introduced. 2. u. |κf (x)| ≤ Lκf |x|. given an auxiliary control law κf 0 ∈ Xf for all Xf ⊆ X . κf (x) ∈ U . f (x. Assumption 3. Xf closed. 0) = 0 αd are K∞ -functions. 49 where is the rst element of The parameters of the controller are the prediction horizon.4 1. Assumption 3. the terminal ingredients are added to provide closed-loop stability as it can be seen in the following section. d1η ∈ D1η .4. tools for analyzing stability of closed-loop min-max MPC systems are provided. Firstly. Xf are such that. 3. while the second one is based on a dual-mode strategy. d1 . there exist a pair of suitable K∞ -functions αVf and βVf such that αVf (|x|) ≤ Vf (x) ≤ βVf (|x|) x ∈ Xf . 0.

N − 1 functions then also κf must are restricted belong to this class. . κf (x). Assumption 3. when semi-feedback controllers are used. 5.e. Min-Max NMPC: an overview on stability Vf (f (x. d1 . In fact. i. . . all d1η ∈ D1η .1) and u(k) = κf (x).5 • Ξ = X MP C • Ω = Xf • α1 = αl • α2 = βVf • α3 = αl • λ 2 = αd Let sup • c1 = N (D2 ) sup • c2 = (D2 ) . for all x ∈ Xf .18) is employed as an ISpS-Lyapunov function.5 If the feedback policies to belong to a particular class of κi (x).4 implies that the closed-loop system formed by (3. is to include the auxiliary control law among the regressors (see the example in Section 3. N ) (3. κf (x). κo [t.4 should hold with κf (x) = 0. and all d2 ∈ D2 . In what follows. . V (x.t+N −1] .6). i = 0. where is function. d1[t. is ISS in Xf (Vf is an ISS-Lyapunov function in Xf ). N ) o o J(x. Assumption 3.50 Chapter 3. d2 ) − Vf (x) ≤ −l(x. This motivates the diculty to guarantee closed-loop stability if optimization is performed with respect to open-loop strategies [Chen et al. d2 ) + (|d2 |). a K∞ - Assumption 3.t+N−1] . On the contrary. d2[t. a natural choice. 1997].t+N−1] . Remark 3. the optimal value of the performance index. d1 .

3. Min-max model predictive control D2 sup D2 ).6 such that In the proof of Theorem 3. it would be preferable that the closedloop system is ISS. ˜ On the contrary. irrespective of the fact that the disturbances may vanish in reality.7 policies As discussed in Remark 2. it is shown that. The previous theorem formulated for the general case of standard minmax MPC states that only ISpS is guaranteed for the resulting closed-loop system. In the following subsections some ingredients are presented. in order to prove the ISpS. is an available feasible solution that guarantees ISpS or ISS. subject to constraints (3.1. so that nominal asymptotic stability is recovered when disturbances are no longer active. the upper bound (3. . this could be a limitation due to (3. Indeed this sequence is such that the value function satises (3. upper bound could be more conservative.N ] .3-3. that make possible to establish ISS. IΩ.N −1] κf ].3.18). However. The only requirement on the possible sub-optimal solution is to be not worst than κ[1.10) in a local region is sucient. 51 The set is such that the set Θ (depending from dened in (3.3. 2007a].2)-(3. respect to d2 with RPIA set X Remark 3. the applicability of a sub- optimal solution of the maximization of the FHCLG is still an open issue [Alamo et al.4.5. Lazar et al. κ where κ[0.N ] ˜ [˜ [0.17). In order to enlarge the set of admissible uncertainty it could be useful to nd an upper bound in a region in [Limon et al. of closed-loop min-max MPC systems. it is not necessary to obtain the In fact the vector of feedback control global optimum solution of the FHCLG in order to guarantee the ISpS (or the ISS) of the closed-loop system. is contained in The main result can now be stated. Raimondo et al.6).13) with function V given by (3. the closed-loop system formed by (3.5. However. 2005.3 Under Assumptions 3. Theorem 3. In fact the uncertainty should be Θ ⊆ Ω.1) and (3. is ISpS with M P C (N ). 2006a.13). κ[1. instead of ISpS.N −1] ˜ is the possible sub-optimal solution obtained at the previous step.11). Ω1 ⊇ Ω as suggested but the However this idea can either enlarge or restrict the set of admissible uncertainty since Ω1 ⊇ Ω Remark 3. 2008]. in the form of assumptions on the type of disturbances or the min-max MPC cost function. when this is the case.

Theorem 3. the following ISpS result (which can be recovered also from Theorem 3.1 Standard min-max with only state dependent uncertainty Consider the case system (3. 2006a]. u) ≥ αl (|x|).8 in Note that Assumption 3. Min-Max NMPC: an overview on stability 3.3-3. Under the above assumption.2 Standard min-max with state independent uncertainty Consider the case system (3.2) (it is known that d1 satis- d2k = 0. which is stated in the following theorem.1) is aected by both uncertainties of the type d1 and of the type d2 . is disturbance independent and where αl is a K∞ -function. when both state dependent and state independent uncertainties are present.1) is aected only by the uncertainty fying (3. Under Assumptions 3. This assumption led to the result published in [Mayne 2001]. that is. Moreover.52 Chapter 3.4. 3. . u) l(x.5. will be considered.4 states that control law u = κf (x) is designed in such a way that the closed-loop system has a stability margin Xf .6 such that The stage cost and l(0.4. the more challenging case. Assumption 3.17).4 [Mayne 2001] Consider that d2 = 0. note that the robustness of the auxiliary control law is translated to the MPC. Next. ∀k ≥ 0).1. with RPIA set X Remark 3.1. A new condition on the stage cost (standard min-max stage cost) is introduced.3) for min-max MPC was obtained in [Limon et al. the min-max MPC controller extends to X M P C (N ) the stability margin provided by the auxiliary control law.2)-(3.1.6). the origin of the closed-loop system formed by (3.1) and (3. 0) = 0 l(x. subject to constraints (3. is robustly asymptotically stable M P C (N ). 3.

using the proof of . In the followings two solutions for solving this problem of standard min-max MPC will be discussed.3. which is attempted in Theorem 3. which will turn out to be sucient for proving ISS of the corresponding min-max MPC closed-loop system. subject to constraints (3. Then.7 The stage cost is composed by the functions q and ld : IR → IR as follows lx : IRn × l(x. d1 .4 can be satised also with ≡ 0. As already mentioned.1. presented previously in Chapter 2.6.1. 3. as it will be shown in Section 3.4 and whose necessity is related to the stage cost considered in the standard min-max MPC optimization problem.17). even if the auxiliary control strategy guarantees ISS. Min-max model predictive control Corollary 3. to the well-known H∞ strategy.3 H∞ strategy The proof of Corollary 3. 3. 2006a] Under Assumptions 3.4. This observation leads to the following new condition on the stage cost. The rst solution.6).5. as the employed control design method prevents that closed-loop asymptotic stability is attained when there are no disturbances active. only ISpS can be established for the min-max MPC closed-loop system. d1 ) − ld (d2 ) and satises Assumption 3. for example.1) and (3. u.1 clearly illustrates that the diculty of proving ISS of min-max MPC. leading.1 53 [Limon et al. The second solution exploits the ISS property of the auxiliary control law via a dual-mode strategy to establish ISS of the dual-mode minmax MPC closed-loop system. this is not a desirable property.3.3. If Assumption 3. u.4 can be satised also with ≡ 0. is related to the terms c1 and c2 that are depending on the K∞ -function dened in Assumption 3.4-3.4. the closed-loop system formed by (3. (3. is ISpS with respect to d2 with RPIA set X Note that. employs a particular design of the stage cost such that Assumption 3. then Assumption 3. IRm × IRp → IR Assumption 3.7 is satised. d2 ) = lx (x.2)M P C (N ).

6). only state independent uncertainties were considered.1. let recall the classical dual-mode strategy.e. if if uDM κM P C (x) κf (x) x ∈ X M P C (N ) \ Xf . .2 with [Magni et al. 3.1) and (3.3. which depends solely on the disturbance signal. 2006a] Under Assumptions 3. the receding horizon controller (3. as established by the following result.7 ≡ 0. 3. ISS of the closed-loop system can be guaranteed. The above result shows that by adding a new term to the stage cost. 2008]. First. In dual-mode MPC.5.4 Dual-mode strategy Sucient conditions for input-to-state stability of nonlinear discrete-time systems in closed-loop with dual-mode min-max MPC controllers were recently developed in [Lazar et al. x ∈ Xf . which relies on the same cost function as the one used in standard min-max MPC.17). i. constraints (3. In this section the results presented in [Lazar et al.54 Chapter 3.4. a dual-mode strategy for guaranteeing ISS of min-max MPC will be presented.19) Next. 3. (3. is ISS with respect to the closed-loop system formed by (3. Corollary 3.4. 3. subject to d2 with RPIA set X M P C (N ).2)-(3. 2008] will be exploited and Theorem 3. the ISS result for dual-mode min-max MPC with standard cost function are presented. Therein.1. Next. It remains to be explored how the modied cost function aects the solvability of the min-max optimization problem and whether standard min-max MPC solvers can still be employed. Min-Max NMPC: an overview on stability Theorem 3. ISS of the resulting min-max MPC closed-loop system can be attained.2 will be used to apply the dual-mode approach to the more general class of uncertainties considered in this chapter.17) is employed outside control law Xf and the auxiliary κf is used inside Xf .

both these methods rely on a specic cost function that must satisfy certain assumptions. is ISS with M P C (N ). However.6). H1 = ∂x x=0 . f1 . subject to constraints (3.4 can be derived by the solution of the H∞ control problem for the linearized system. For convenience. Therefore.21) .3-3. the computation of suitable cost functions is equally important for synthesizing min-max MPC schemes with an a priori ISS guarantee.5.1) and (3. f3 and h1 are C 2 functions with f1 (0) = 0 and h1 (0) = 0. Given a square n × n matrix P .1.20) h1 (xk ) uk w = [d1 d2 ] . respect to d2 with RPIA set X So far. 3. ered.5 55 Under Assumptions 3.19).2)-(3. The auxiliary control law Theorem 3. 2003]) also to the presence of state independent disturbances. two methods for establishing ISS of min-max MPC have been presented. f2 . system In this respect.3.5 The auxiliary control law In this section it is shown that. let represent the corresponding discrete-time linearized system as xk+1 = F1 xk + F2 uk + F3 wk zk = where H1 x k uk ∂h1 F1 = ∂f1 ∂x x=0 . A solution that applies to nonlinear systems ane in control is presented in the next section. 3.5. consider the xk+1 = f1 (xk ) + f2 (xk )uk + f3 (xk )wk zk = where (3. dene also the symmetric matrix R = R(P ) = R11 R12 R21 R22 (3. F3 = f3 (0). a nonlinear control law This section extends the results obtained in [Magni et al. and a positive constant γ . F2 = f2 (0). if nonlinear input ane systems are consid- u = κ∗ (x) satisfying Assumption 3. the closed-loop system formed by (3.

0 P Kdx ≥ 0 and Kdu ≥ and that there exists a positive denite matrix (ii) (iii) such that R22 < 0 −P + F1 P F1 + H1 H1 − F1 P u= F2 F3 R−1 F2 F3 P F1 < 0.56 Chapter 3. Suppose that with |d1 |2 = |d1η η(|(x. u.1 is satised. d2 ) = |zl |2 hl (x) .4 with stage cost (a) or (b) l(x. Then. Proposition 3. u. there exist sets D2 the control law D1 (x) and D2 such that for all ∗ (x) where κ = −R(x)−1 f2 (x) f3 (x) ¯ w ∈ Wne = D1 (x)× κ∗ (x) ξ ∗ (x) with P f1 (x) R(x) = f2 (x) P f2 (x) + I f2 (x) P f3 (x) f3 (x) P f2 (x) f3 (x) P f3 (x) − γ 2 I = r11 (x) r12 (x) r21 (x) r22 (x) satises Assumption 3. b c bL2 f and (s) γ 2 |s|2 κ with zl = such that γ 2 Kdu .1 (i) Suppose that Assumptions 3. d1 . Min-Max NMPC: an overview on stability where R11 = F2 P F2 + I R12 = R21 = F2 P F3 R22 = F3 P F3 − γ 2 I and the quadratic function Vf (x) = x P x. . u)|)|2 ≤ Kdx |x|2 + Kdu |u|2 . a γ 2 Kdx . d2 ) = |z|2 − γ 2 |w|2 and ≡0 l(x. d1 . where hl is u h1 (x) h1 (x) ≥ hl (x) hl (x) + a|x|2 + c|x|2 .

The parameters of the system are M = 1 kg . α is a nite positive constant. Finally a damper with damping factor hd aects the system in a resistive way.33 N . d1 .3: Cart and spring-damper example. time state space nonlinear model x1 (t) = x2 (t) ˙ k0 x2 (t) = − M e−x1 (t) x1 (t) − ˙ where hd M x2 (t) + u(t) M + d2 (t) M x2 is the carriage velocity. u. The model of the system is given by the following continuous- Figure 3. d1 . 3. Remark 3. k0 = 0. This carriage (see Figure 3. while the damping factor in not well known and is given m .4 with both the stage cost l(x. where x1 is the displacement of the carriage from the equilibrium position associated with the external force u = 0 and the external disturbance force (wind force) d2 = 0. the MPC law introduced in the chapter is applied to a cart with mass M moving on a plane (the model is the same of the paper [Magni et al.3) is attached to the wall via a spring with elastic constant k given by k = k0 e−x1 .9 P can be computed by solving a discrete-time H∞ algebraic Riccati equation.3. 2003]). Example 57 and Xf where x : x Px ≤ α ⊆ X.6. u. It is important to underline that the proposed auxiliary control law satises Assumption 3.6 Example In this section. d2 ) = |zl |2 . d2 ) = |z|2 − γ 2 |w|2 and l(x.

1 N s and |d1η | ≤ 0.01 and Kdu = 0. d2 ) = |z|2 − γ 2 |w|2 and l(x.4.58 Chapter 3.4 s is given by x1k+1 = x1k + Tc x2k ¯ k0 x2 = −Tc M e−x1k x1k + x2k − Tc hd x2k M k+1 d1η d2 +Tc uk − Tc Mk x2 (k) + Tc Mk M system with which is a discrete-time nonlinear system. d1 .1 Kdx = 0.5 with H1 0 q2 u x1 x2 z= h1 (x) u = q1 0 . The system can be rewritten as x1k+1 x2k+1 = 1 k0 −Tc M e−x1k 0 − Tc M F3 Tc ¯ 1 − Tc hd M f1 (xk ) x1k x2k + 0 Tc M F2 uk + 0 Tc M d1k d2k wk with d1k = d1ηk x2k . q1l 0 0 q2l u x1 x2 with q1l = 1 and q2l = 1 and The auxiliary control law is obtained as described in Section 3.65 m.2 ≤ d2 ≤ 0. u) = |zl |2 where H where L zl = hl (x) u = γ = 3.5 N . where hd = 1. Min-Max NMPC: an overview on stability by ¯ ¯ hd = hd + d1η .1. An Euler approximation of the sampling time Tc = 0. m −0. The system can be rewritten as x1 (t) = x2 (t) ˙ k0 x2 (t) = − M e−x1 (t) x1 (t) − ˙ ¯ hd M x2 (t) Wind force is limited: + u(t) M − d1η (t) M x2 (t) + d2 (t) M The state and control variables have to satisfy the following constraints |u| ≤ 4. u. |x1 | ≤ 2. Disturbance d1k satises point (i) of Proposition 3. Let choose l(x.

Moreover.09. in the neighborhood of the origin. as required at point b) of Proposition h1 (x) h1 (x) ≥ hl (x) hl (x) + a|x|2 + c|x|2 with a = γ 2 Kdx = 0.3708 3. other hand. q1 = 1.7}. when the disturbance vanishes. For t ≥ 1.0814 3. the auxiliary control law is a good approximation of the H∞ strategy.7 show the time evolution of d1η and d2 . Figure 3.6. The auxiliary control law satises Assumption 3. standard min-max strategy only guarantees ISpS. c = bL2 f = 0.2998 H∞ algebraic Riccati equation is computed solving a discrete time P = F1 P F1 + H2 H2 − F1 P with F2 F3 R−1 F2 F3 P F1 H2 H2 = 1. The policies κi (x) 2 2 are functions of the form κi (x) = αi κf (x) + βi (x1 + x2 ) + γi .3708 4. Figure 3. starting from Xf x1 (0) = 0.09x2 + 0.1204]f1 (x) P = 7. H∞ and dual mode strategies guarantee On the ISS: cart position goes to zero when the disturbance vanishes.2H1 H1 in order to satisfy inequality inequality (iii) of Proposition 3. 1 2 1 2 1 2 The auxiliary control law is given by κf (x) = −[1 0 0]R−1 where F2 F3 P f1 (x) = −[0. In fact κ 2 2 2 2 q1 x2 + q2 x2 ≥ q1l x2 + q2l x2 + ax2 + ax2 1 2 1 2 1 2 1.1. The length of horizon is N = 4. The terminal penalty is given by Vf = x P x. Example 59 with 3.1541m.1.5 show the time evolution of position and velocity of the cart.1.6 shows the control sequence.1 and q2 = 1. s Note that.21x2 + 1. where Lκf = 2. Region Xf has been obtained numerically.21x2 ≥ x2 + x2 + 0. Note that.5m and x2 (0) = 0 m . note that that H∞ and dual mode performances are comparable since. cart position does not tend to the origin but to x1 = 0. for the stages {x : x P x ≤ 4.09x2 .4 and cost chosen. Figure 3.1.3. In fact.4. . b = γ 2 Kdu = 0. in the region 3.8783 1. Matrix P satises (ii) of Proposition 3.6s the signal d2 is equal to zero.

05 0 −0. Min-Max NMPC: an overview on stability 0.25 Figure 3.60 Chapter 3.3 0 1 2 3 4 time 5 6 7 8 −0.05 velocity −0. 0.1 −0.2 0.4: Time evolution of cart position.15 −0.6 Standard min−max H∞ strategy Dual mode strategy 0.3 0. .1 0 1 2 3 4 time 5 6 7 8 Figure 3.4 position 0.1 0 −0.5: Time evolution of cart velocity.2 Standard min−max H∞ strategy Dual mode strategy −0.5 0.

2 Input −0.6: Time evolution of the control.2 0.1 0.6.15 d2 0 1 2 3 4 time 5 6 7 8 Figure 3.15 Noise 0.1 0 −0. 0. .4 −0.3.1 −0.1 −0.5 −0.6 −0.7 Standard min−max H∞ strategy Dual mode strategy 0 1 2 3 4 time 5 6 7 8 Figure 3. Example 61 0.25 0.05 −0.3 d1η 0.7: Time evolution of uncertainty d1η and disturbance d2 .3 −0.05 0 −0.

7).7 Conclusions In this chapter a unied framework for the synthesis of min-max MPC control algorithms has been provided.12) is a RPI set for system (3. all d1 ∈ D1 (x). The ISpS or ISS property of such algorithms is analyzed with respect to a general class of disturbances that considers both state dependent and state independent disturbances. Then ∆V (x) ≤ −α4 (V (x)) + ζ(c1 ) + c2 + λ2 (|d2 |) = −α4 (V (x)) + λ2 (|d2 |) + c3 ≤ −α4 (V (x)) + λ2 (||d2 ||) + c3 . and all d2 ∈ D2 .62 Chapter 3. 3. t such that xt ∈ Θd2 .8 Appendix Proof of Theorem 3. The relevance of the adopted stage cost to achieve ISS claries the dierence between some of the results appeared in the literature. Moreover (see [Limon et al. this implies ρ ◦ α4 (V (xt )) ≤ λ(||d2 ||) + c3 . where c3 c2 + ζ(c1 ). 2006a]): 2 rst. Therefore V (x) ≤ α2 (|x| + c1 ) and hence |x| + c1 ≥ α−1 (V (x)). ∀x ∈ Ω. Min-Max NMPC: an overview on stability 3. on an The H∞ cost and on a dual mode approach are compared. Θd2 dened in (3. Without loss of generality. for all x ∈ Ξ. assume that (id − α4 ) is a K∞ -function (see Lemma Assume now that there exists a nite time . it is going to be shown that α3 (|x|) + ζ(c1 ) ≥ α3 (|x| + c1 ) ≥ α4 (V (x)) where (3. algorithms based on a standard stage cost.2: Step 1: The proof will be carried out in four steps.22) α4 α3 ◦ α−1 2 is a K∞ -function. From the denition of α2 (s) it follows that α2 (|x|) + c1 ≤ α2 (|x| + c1 ). Then V (xt ) ≤ b(λ2 (||d2 ||) + c3 ).

By induction one can show that for all j ∈ Z≥0 . one has ρ(α3 (|x|) + ζ(c1 )) > λ2 (||d2 ||) + c3 . d2 ) ∈ Θd2 for all k ≥ t. d1t . d2k0 )) ≤ −(id − ρ) ◦ α4 (b(λ2 (||d2 ||) + c3 )) +b(λ2 (||d2 ||) + c3 ) ≤ b(λ2 (||d2 ||) + c3 ).1 [Jiang & Wang 2001]). if x ∈ Ω \ Θd2 . d2 )) ≤ b(λ2 (||d2 ||) + c3 ) ¯ x(t + j.7). Then V (F (xt . Hence Θd2 is ¯ a RPI set for system (3. Firstly. x. it follows that ρ ◦ α4 (b(s)) = s K∞ -function. d1k0 . starting from Ξ \ Θd2 .23) ∆V (x) < 0. ρ ◦ α4 (V (x)) > λ2 (||d2 ||) + c3 . Step 2: now. then the state tends asymptotically to Θd2 .22). it follows that and. On the other hand. ∀x ∈ Ω \ Θd2 . d1 . (id − ρ) is a K∞ -function. d1 . .3. x. From the inequality (3. Appendix 63 B. owing to the V (F (xk0 . From the denition of fact that (id − ρ) is a b. Hence α3 (|x|) > λ2 (||d2 ||) + c2 that means (3. d2t )) ≤ (id − α4 )(V (xt )) + λ2 (||d2 ||) + c3 ≤ (id − α4 )(b(λ2 (||d2 ||) + c3 )) + λ2 (||d2 ||) + c3 = −(id − ρ) ◦ α4 (b(λ2 (||d2 ||) + c3 )) +b(λ2 (||d2 ||) + c3 ) − ρ ◦ α4 (b(λ2 (||d2 ||) + c3 )) +λ2 (||d2 ||) + c3 . it is shown that. that is V (x(t + j. ∀s > 0 then α3 (|x|) + ζ(c1 ) > ρ(α3 (|x|) + ζ(c1 )) > λ2 (||d2 ||) + ζ(c1 ) + c2 . hence id(s) > ρ(s).8.

¯ Ξ. ∀x1 ∈ Ξ \ Ω. d2 )|Θd = 0. x. starting from ∆V (x(T1 . ∀x ∈ Ξ \ Ω c so that. d1 . d2 ) ∈ Ω. d1 . d1 . let R(e) {x : V (x) ≤ e = maxR(e)⊆Ω e}. such that ∀ε > 0. starting from time and to time. Then from (3.2). if x(T1 . d1 . d2 ))) ¯ −ρ ◦ α4 (V (x(T1 . the state arrives in Θd2 in a nite ¯ Therefore. Then. x. Hence limk→∞ |x(k. d1 . x. ¯ If in particular Ξ. x. d2 ) ∈ Θd2 . d1 . ¯ / ∆V (x(T1 . d1 . the state will reach the region Ω in a nite time. Min-Max NMPC: an overview on stability Θd2 (see point 4 of Denition 3. ¯ Then ∆V (x) < −¯θ . d1 . ¯ 2 Otherwise.23) it follows that ¯ Moreover. x. there exists a c2 > 0 such that. for x(T1 . the region Θd2 is achieved in a nite ¯ time. ¯ Note that if ||d2 || = 0 or c3 > 0.24) V (x(T2 . d2 )|) ¯ where the last step is obtained using (3. x. there exists x2 ∈ Ω \ Θd2 such that ¯ α3 (|x2 |) ≤ α3 (|x1 |) − cθ . x. x. d1 . d1 . Since Θd2 is a RPI set. d1 . x. there is |x(T1 . d2 )) ≤ ε + b(λ2 (||d2 ||) + c3 ). d1 . x. it is true that limk→∞ |x(k. x. there exists T1 such that x(T1 .5 in [Jiang & Wang 2001]. d1 . the state will arrive close to Θd2 in a nite Θd2 asymptotically. d2 ) ∈ Θd2 . d2 ))) + λ2 (||d2 ||) + c3 ¯ ≤ −(id − ρ) ◦ α4 (V (x(T1 . using the same steps of the proof of Lemma 3. x. that means. d2 ) ∈ Θd2 . d2 ) ∈ Θd2 . considering that −1 |x| ≤ α1 (V (x)). d2 )) ≤ ¯ = −α4 (V (x(T1 . x. ∃T2 (ε ) ≥ T1 (3. It ¯ given is a RPI set.64 Chapter 3. d1 . . d2 )) > λ2 (||d2 ||)+ ¯ / ¯ c3 and Therefore. d1 . x(T1 . ρ◦α4 (V (x(T1 . Let is clear that Ψ ⊇ Θd2 and that of V (x) is known in Ψ ⊆ Ω then. d2 ))) ¯ ≤ −(id − ρ) ◦ α4 ◦ α1 (|x(T1 . there exists a c > 0 such that ˜ Θsup = c. there exists cθ > 0 such that for all x1 ∈ Ξ \ Ω. d2 )) < −c2 .9). Then. by denition of −α3 (|x1 |) + cθ ≤ −α3 (|x2 |) < −λ2 (|d2 |) − c2 . d2 ))) + λ2 (||d2 ||) + c3 ¯ −(id − ρ) ◦ α4 (V (x(T1 . Step 3: Ψ Ψ e ∈ IR≥0 . d1 . for x(T1 . x. d2 )| > c. ˜ ¯ / ¯ ˜ d2 Hence. x. d1 . d2 )|Θd2 = 0. Since the upper bound {x : V (x) ≤ e}. x. x.

8. Lyapunov function in X M P C (N ). it is LpS and UpAG in Ψ. 2006a].1. the lower bound is easily obtained o o V (x. consider the following policy vector . N ) = J(x. x. Step 4: nally it is shown that system (3. X M P C (N ).1) and (3. First. κ[t. Appendix 65 that also hold for discontinuous systems. the robust invariance of tion 3. dened in 3. d2 )) ≤ ε + b(λ2 (||d2 ||) + c3 ) ¯ ¯ hence −1 |x(T2 .t+N −1] . Finally. d1 . N ) ≥ κ[t.t+N −1] . In order to derive the upper bound. x. Ξ. d1 .3. d1 . it follows that −1 −1 −1 |x(T2 . then it is ISpS in is an ISpS-Lyapunov function for the closed-loop system (3. see [Limon et al. Ξ imply that the system is ISpS in Ξ with respect to d2 .2. In view of Theorem 3. θ1 (s1 + s2 ) ≤ θ1 (2s1 ) + θ1 (2s2 ).t+N −1] at time t. if system admits an ISpS- X M P C (N ). Now. ¯ −1 −1 α1 (2b(2λ2 (||d2 ||))) and c ε α1 (2ε ) and γ(||d2 ||) −1 α1 (2b(2c3 )).t+N ] ¯ time t + 1 starting X M P C (N ) is easily derived from Assumpκf (xt+N )] as admissible policy o from the optimal sequence κ [t. N ) ≥ l(x.4 by taking vector at κ[t+1. d2 )| ≤ α1 (2ε ) + α1 (2b(2λ2 (||d2 ||))) + α1 (2b(2c3 )). the regional ISpS in Ψ is obtained. x. In the following it will be shown that the function V (x. letting since the system is regional ISpS in LpS with UpAG in Ψ.t+N −1|t] Then.25) x ∈ X M P C (N ). κo [t.t+N −1] .7) is regional ISpS in Using (3. N ). d1[t. x.17) in Proof of Theorem 3. d2 )| ≤ α1 (ε + b(λ2 (||d2 ||) + c3 )).24) and (3.3. the UpAG property in Ξ is proven. 0. given a K∞ -function θ1 .18. 0.t+N −1] . 0. 0) ≥ αl (|x|) for all (3. using Assumption 3. d2[ t. ¯ Noting that. d1 .9) there is α1 (|x(T2 . [κo [t+1.3: by Theorem 3. κ0 (x).t+N −1] min J(x. d2 )|) ≤ V (x(T2 .

d2 ) + (D2 ). N + 1) ≤ ≤ d1η ∈MD1η .4. d1 . N ) − V (x. d2 ). uk . uk . see the analogous proof of Theorem 2. Min-Max NMPC: an overview on stability [κo [t. N − 1) + (D2 ) ≤ . Hence. . d1k .3. d1[t. d1k . d1[t. d2[t.d2 ∈MD2 max max J(x.t+N −1] κf (xt+N )] as admissible t with horizon N + 1. κ[t. ut+N . 0) + N (D2 ) sup sup = Vf (x) + N (D2 ) ≤ βVf (|x|) + N (D2 ) for all (3. d2[t.t+N ] . V (f (x. d1 . κ[t.t+N ] ˜ at time Chapter 3. From the monotonicity property (3.26) x ∈ X M P C (N ). N + 1) ≤ ˜ t+N −1 l(xk . d2 ). the upper bound is obtained sup sup V (x.t+N ] .t+N ] . N ) + (D2 ) which holds for all (3. κM P C (x). N − 1) − V (x.26) and by sub-optimality there is For details.4 J(x. . κM P C (x).27) x ∈ Xf . d1 . and all d2 ∈ MD2 . N ) sup ≤ −l(x.t+N ] . d2t+N ). all d1η ∈ MD1η . d2k ) + Vf (xt+N ) k=t + (|d2 |) sup = V (x. d2[t.t+N ] .t+N ] . d2k ) + Vf (xt+N ) + (|d2 |) k=t which implies V (x. d1[t. κM P C (x).t+N ] . N ) ≤ V (f (x. Therefore. by applying . N ) ≤ V (x. Then policy vector for the FHCLG J(x. using Assumption 3. d1k . κ[t.d2 ∈MD2 l(xk .t+N ] . N + 1) = ˜ t+N −1 l(xk .t+N ] . d2k ) + Vf (xt+N ) − Vf (xt+N ) + Vf (xt+N +1 ) k=t +l(xt+N . N + 1) ˜ t+N −1 d1η ∈MD1η . uk . d1t+N . In view of Assumption 3.66 κ[t. ≤ V (x.

17) X M P C (N ). d2 ). In particular. Hence the closed-loop system with the auxiliary control law is ISS in is obtained.19) is ISS in Xf . In Corallary 3. is ISpS with respect to d2 with RPIA set X M P C (N ).1) and (3.5. Finally.4 for a disturbance d2 dierent from zero. the closed-loop system (3. a standard stage cost l(x. X M P C (N ) and hence. Xf . As- sumptions 3.1)-(3. by Theorem 3. it will be shown that the .3. This fact. N ) − V (x.2 is satised by the inequality (3. One of the key steps in the proof of Theorem 3. it is shown that condition 3 in Denition 3. by As- sumption 3.3.28).4.5 guarantee that the closed-loop system (3.3-3.6. Only ISpS can be proven because of term sup (D2 ) derived by term (|d2 |) in point 5 of Assumption 3.17) in Theorem 3. κM P C (x).1. the bound on the decrease of is obtained sup V (f (x. Following the steps of the proof of Theorem 3.5: is ISpS in As shown in the proof of Theorem 3. all d1η ∈ MD1η . using Assumption 3.4. term (|d2 |) must be dierent from zero.1. Since Xf is achieved in a M P C (N ) nite time and system satises UAG property in Xf . In order to guarantee the satisfaction of Assumption 3. and all d2 ∈ MD2 .1. u) is considered.17). Appendix V 67 Assumptions 3. subject to constraints (3.2. However.1 is derived by proof of Theorem 3. by Therefore.2 is satised.6).27).1)-(3. Proof of Corollory 3. V (x) is an ISpS-Lyapunov function for the closed-loop system (3.2)-(3. the closed-loop system formed by (3. X M P C (N ) since UAG in X M P C (N ) and LS are equivalent to ISS in X M P C (N ). considering Assumption 3. point 5 of Assumption 3.3.1: Proof of Corollary 3. αd ≡ 0.25). note that in this case.26).2.28).4 and monotonicity property (3. The necessity of this term is related to the particular stage cost considered in the optimization problem.3 is to show that condition 3 in Denition 3. 3. by (3. The auxilor when it iary control law is used when the state reaches the region starts in Xf Vf (x) is an ISS-Lyapunov function in Xf . (3. Proof of Theorem 3.3. Proof of Proposition 3. (3.13).1) and (3.1: In the following. N ) ≤ −αl (|x|) + αd (|d2 |) + (D2 ) (3.8. leads to a less conservative estimation of the region Θ dened in (3. it can be proven that region Xf is achieved in a nite time. d1 . ISS in Xf is equivalent to UAG in Xf and LS.3.28) for all x ∈ X M P C (N ). UAG in X By Assumption 3.

u.4. κ∗ (x). Then (3.4 is x belonging to Ω0 = {x : |x| ≤ r0 } ⊂ X . there exists a positive constant satised for all First.30) H(x. computing H(x. By the Taylor expansion Theorem (note that the rst order term is evalu- . w) = = (f1 (x) + f2 (x)u + f3 (x)w) P (f1 (x) + f2 (x)u + f3 (x)w) −x P x + h1 (x) h1 (x) + u u − γ 2 w w f1 (x) P f1 (x) − x P x + h1 (x) h1 (x) +u +2 f2 (x) P f2 (x) + I u + w u w f2 (x) P f1 (x) f3 (x) P f1 (x) u w f3 (x) P f3 (x) − γ 2 w + 2u F2 (x) P F3 (x)w = f1 (x) P f1 (x) − x P x + h1 (x) h1 (x) + u w R(x) for +2 and u w f2 (x) P f1 (x) f3 (x) P f1 (x) and. Min-Max NMPC: an overview on stability proposed auxiliary control law satises Assumption 3. w) = Vf (f1 (x) + f2 (x)u + f3 (x)w) − Vf (x) + |z|2 − γ 2 |w|2 . note that r0 such that point 2 of Assumption 3. u. w) u = κ∗ (x) w = ξ ∗ (x). for all x ∈ Ω1 = {x : |x| ≤ r} ⊂ Ω0 . κ∗ (x). ξ ∗ (x)) ≤ −ε|x|2 . ξ ∗ (x)) = f1 (x) P f1 (x) − x P x + h1 (x) h1 (x) + · κ (x) ξ ∗ (x) ∗ ξ ∗ (x) R(x) · −2 κ∗ (x) ξ ∗ (x) R(x) = f1 (x) P f1 (x) − x P x + h1 (x) h1 (x) − f1 (x) P f2 (x) f1 (x) P f3 (x) R(x)−1 From (iii) it follows that there exist positive constants ε.29) H(x. κ∗ (x) κ (x) ξ ∗ (x) f2 (x) P f1 (x) f3 (x) P f1 (x) ∗ H(x. Dene H(x.68 Chapter 3. r such that (3. u.

u. d2 ) = |z|2 − γ 2 |w|2 and for all satised. d1 .29). ∀x ∈ Xne . ∀w ∈ Wne . there exist an open neighbor- H(x. 2 r22 (x) < 0. d1 . κ∗ (x).4 is as dened in point ≡ 0. . w) = H(x. ξ ∗ (x)) ≤ −ε|x|2 for all x ∈ Ω2 = {x : |x| ≤ r2 } ⊂ Ω1 such that and for all w ∈ Wne . w∗ ) + 1 2 u − κ∗ (x) w − ξ ∗ (x) R(x) u − κ∗ (x) w − ξ ∗ (x) of Assumption (ii) implies that there exists a neighborhood that for all then Xne x ∈ Xne . Let choose some β>0 Ωβ = {x : Vf (x) ≤ β} ⊂ Ω2 . κ∗ (x). if H∞ strategy is used with Assumption 3. κ∗ (x). given such that Xne . d2 ) = |zl |2 b) of Proposition 3. w) ≤ H(x. there exists a positive constant r2 such that (3. From (3. κ∗ (x). it Wne of w = 0 follows that.31). w) = H(x. u. κ∗ (x). If the system is controlled by x = 0 such u = κ∗ (x) H(x.8.1 (standard min-max stage cost) Vf (f1 (x) + f2 (x)κ∗ (x) + f3 (x)w) ≤ Vf (x) − |z|2 + γ 2 |w|2 = Vf (x) − h1 (x) h1 (x) − u u +γ 2 d1 (x)d1 (x) + γ 2 d2 d2 . Appendix (u.32) follows that (3. κ∗ (x).31) H(x. κ∗ (x). l(x.32) Vf (f1 (x) + f2 (x)κ∗ (x) + f3 (x)w) ≤ Vf (x) − |z|2 + γ 2 |w|2 − ε|x|2 ≤ Vf (x) − |z|2 + γ 2 |w|2 x ∈ Ωβ and all w ∈ Wne . (3. In view of (3.30). ξ ∗ (x)). w) = (κ∗ (x). ξ ∗ (x)) + Since hoods 1 (w − ξ ∗ (x)) r22 (x) (w − ξ ∗ (x)) . w) ≤ H(x. point 5 of l(x. (3.3. Consider now the case of Hence. u. ξ ∗ (x)) >2 69 ated in and terms of order are null) H(x. r22 (x) < 0.

and all w ∈ Wne where (s) γ 2 |s|2 . dening Xf as Xf there is {x : Vf (x) ≤ α}. Xf is a RPI set.4 is satised for the standard min-max stage cost too. 3. Θ(D2 ) is a RPI set. Since Θ(D2 ) ∝ sup ¯ D2 . 4. using point b) of Proposition 3. κ Then.4 is satised. Min-Max NMPC: an overview on stability Using point (i) of Proposition 3. It is clear that The invariance of the closed-loop system (with the auxiliary control law) in Xf ends the proof that Assumption 3.1 Vf (f1 (x) + f2 (x)κ∗ (x) + f3 (x)w) ≤ Vf (x) − h1 (x) h1 (x) − u u +γ 2 (Kdx |x|2 + Kdu |u|2 ) + γ 2 d2 d2 ≤ Vf (x) − h1 (x) h1 (x) − u u + a|x|2 +b|u|2 + γ 2 d2 d2 with a γ 2 Kdx and b γ 2 Kdu . let consider the sup sup proof of Theorem 3.4 is locally satised Vf (f1 (x) + f2 (x)κ∗ (x) + f3 (x)w) ≤ Vf (x) − h1 (x) h1 (x) − u u +a|x|2 + c|x|2 + γ 2 d2 d2 with c bL2 f . . Hence point 5 of Assumption 3. u) + (|d2 |) for all x ∈ Ωβ .1 Vf (f1 (x) + f2 (x)κ∗ (x) + f3 (x)w) ≤ Vf (x) − hl (x) hl (x) − u u + γ 2 |d2 |2 ≤ Vf (x) − hl (x) hl (x) − u u + (|d2 |) = Vf (x) − l(x. By step 1.70 Chapter 3.2. Point 4 Vf (x) = x P x. In order to verify that Assumption 3. Considering that point 2 of Assumption 3. there exists a set Wne = D1 (x) × D2 ⊂ Wne and a positive constant In order to prove point α such that. Point is obviously is such that αVf (|x|) where λmin (P )||x||2 ≤ x P x ≤ λmax (P )||x||2 βVf (|x|) λmin (P ) and λmax (P ) are minimum and maximum eigenvalues of P . Θ ⊂ Xf ⊂ Ωβ . with P positive denite matrix. the robust invariance of Xf . 3. satised since 1 is obtained if Xf ⊆ Ωβ ⊂ X .4 is satised. it remains to prove points 1.

4 4. 4. . . . . .6 4. . . .2 4. . .5 4. . . . . . . . The controller chooses the input state xk so as to ensure constraint satisfaction along the predicted trajec- tory of the plant for any possible uncertainty. . . . . . . . Guaranteed bound of the max function . Example . . . . .7 4. . min-max model predictive control is based on the solution of a nite-horizon game. . . . . . .1 4. . . . minimizing at the same time the worst case performance index of the predicted evolution of the system. Conclusions . . . 71 72 73 76 77 79 81 82 Problem statement Min-Max Nonlinear Model Predictive Control .3 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8 Introduction . . . . . . . . . . .1 Introduction As discussed in Chapter 3. . . .Chapter 4 Min-Max Nonlinear Model Predictive Control: a min formulation with guaranteed robust stability Contents 4. . . . . . . . . . . . where player (the controller) and u is the input of the minimizing uk as a function of the current w is the input of the maximizing player (the nature). . . . . . . . . . . . . . . . . . . . . . . . . Appendix . . . . . Proposed Min formulation: stability . . . . . . . . .

Min-Max NMPC: a min formulation with guaranteed robust stability The calculation of the solution of the optimization problem results to be an NP-hard problem [Blondel & Tsitsiklis 2000]. uk ∈ IRm is the current control vector. even in the case that the cost is a convex function [Scokaert & Mayne 1998] with respect to the future sequence of disturbances. uk . Lee & Yu 1997. For simplicity of notation. Alamo et al. The computational burden of this solution is much lower but at the expense of a potential minor loss of performance.1 1. f (0.2 Problem statement Assume that the plant to be controlled is described by discrete-time nonlinear dynamic model xk+1 = f (xk . This has motivated an increasing eort to nd novel approaches to the min-max MPC which maintain its desirable properties with a more tractable optimization problem associated. i. The system is supposed to fulll the following assumption. the proposed predictive control is shown to inherit the convergence and the domain of attraction of the standard min-max strategy. This is typically done by using an upper bound function of the min-max cost function [Kothare et al. 1996. 4. Lu & Arkun 2000] or by using a relaxed maximization problem aimed to make the problem tractable [Alamo et al. This choice produces a solution that does not dier much from the original min-max problem. 0. Assumption 4. To this aim. the max stage is replaced by a simple suitable choice of an uncertain realization. . it is assumed that the origin is an equilibrium point. This fact has limited its applications to a narrow eld in spite of its benets. wk ). 0) = 0. Moreover. 2005. The goal of this chapter is to propose a relaxed min-max predictive controller for constrained nonlinear systems.e. k ≥ 0 where (4. 2007].72 Chapter 4.1) xk ∈ IRn is the state. wk ∈ IRp is the disturbance term.

Hence.3) (4. The state of the plant xk can be measured at each sample time. . .3. a practical solution.2) ε > 0. compromise between the simplicity of the open-loop min-max strategy and the performance advantages of the closed-loop one. zf ) = κf (x). 4. yielding to an optimization problem similar to the open-loop case one. in the interval u = κf (x) will be used. able of each control law is its set of dening parameters. zt+j ). In this case. The functions [t+Nc . In the following. 3.4) X and U are compact sets.e. Thus. the decision vari- but forcing a given structure of the control law. that is.3 Min-Max Nonlinear Model Predictive Control As discussed in Chapter 3. Nc − 1]. at the end of the control horizon. control policies are considered as decision variables. an auxiliary state-feedback control law κ should be such that the auxiliary zf such that control law is a particular case of the possible feedback control strategies. Min-Max Nonlinear Model Predictive Control w 73 2. κ(xt+j .4. will be considered. a sequence belonging to a nite-dimensional space. both containing the origin as an interior point. zt+Nc −1 ]. . . at each instant t. In the follow- Nc and Np respectively with Nc ≤ Np . 4. the optimal min-max problem can be stated. is the so-called semi-feedback min-max formulation. The disturbance is such that w ∈ W = {w ∈ IRp : |w|∞ ≤ ε} where (4. will be used as feedback control strategies [Fontes & Magni 2003]. t+Np −1]. i.t+Nc −1] [zt . functions of x and parameters z ∈ IRq . j ∈ [0. ing. the controller will have only to choose z[t. . Moreover. dierent control and prediction horizon. The state and the control variables are restricted to fulll the following constraints x∈X u∈U where (4. there exists κ(x. This means that.

for each pos- 3.t+Nc −1] and maximizing with respect to w[t. uk ) + Vf (xt+Np ) (4. the Finite Horizon the cost function Closed-Loop Game (F HCLG) problem consists in minimizing.t+Nc −1] . Min-Max NMPC: a min formulation with guaranteed robust stability Consider system (4.1 (FHCLG) the positive integers the terminal set to xt = x. w[t. with respect z[t. t + Nc − 1]. t + Nc − 1] κf (xk ). the constraints (4. zk ). be a solution of the FHCLG. t + Np − 1] k ∈ [t.t+Np −1] ∈ MW κf (xk ) Letting κ(xk . k ∈ [t + Nc .t+Nc −1] z[t. the same constraint has to be satised between the end of the control horizon and the end of the prediction horizon.2)-(4.t+Nc −1] .74 Chapter 4. k ∈ [t. k ∈ [t + Nc . the stage cost l.4). the control signal l(xk . t + Np − 1] xt+Np ∈ Xf . that means. according to o zo [t. the terminal state constraints Note that the parameters sible sequence w[t. have to be such that. and the auxiliary control law κf . The stage cost denes the performance index to optimize and must satisfy the following assumption. z0 ) where (4. the state dynamics (4.6) o z0 is the rst element of zo [t. t + Np − 1].5) uk = κ(xk . Moreover.t+Nc −1] . z[t.1) 2. the terminal penalty Vf Xf .1) with Denition 4. w[t.t+Np −1] the RH paradigm.t+Np −1] ) x k=t subject to: 1. zf ) ∈ U. zk ) ∈ U. 4. ¯ Given Nc and Np .t+Np −1] t+Np −1 J(¯.t+Nc −1] ∈ MW κ(xk . for all w[t. k ∈ [t. the feedback control law setting u = κM P C (x) is obtained by o κM P C (x) = κ(x. .

t+Np −1] o and z zo [t. where it is shown how to compute a nonlinear auxiliary control law based on the solution of a suitable lem for the linearized system under control. 2003]. In order to simplify the notation. κf (x). z. LDIs (Linear Dierential In- clusions) around the origin and by using standard linear robust strategy A solution for ane systems has been proposed in Chapter 3 and [Magni et al. αVf < βVf for all and x ∈ Xf . the termi- Remark 4. Xf are such that. for all x ∈ Xf and all w ∈ W .2 l(x. given an auxiliary control law κf 0 ∈ Xf Xf ⊆ X . ·) is such that l(0. z) maxw∈MW J(x. Assumption 4. where is a K∞ -function. Min-Max Nonlinear Model Predictive Control Assumption 4. w) ∈ Xf . In order to derive the main stability and performance properties associated to the solution of FHCLG. The design parameters Vf . 2. 0) = 0 and l(x. where the minimization with respect z is carried out under the robust constraints of the FHCLG problem.t+Nc −1] .t+Nc −1] .3. for example. w w[t. u) ≥ u ∈ U. In the following let denote ¯ J(x. The computation of the auxiliary control law.3 1. all w∈W and 4. w) − Vf (x) ≤ −l(x. there exist a pair of suitable K∞ -functions αVf βVf such that αVf (|x|) ≤ Vf (x) ≤ βVf (|x|) x ∈ Xf 5. could be obtained by.4. 3. Moreover. let denote with FHCLG problem to V (x) V (x) the cost associated with the solution of the ¯ minz J(x. Xf κf (x) ∈ U . Vf (f (x. [Alamo et al. l(·.1 nal penalty and the terminal inequality constraint satisfying Assumption 4. for all x ∈ Xf for all f (x. 0). let denote H∞ prob- z z[t. κf (x).3. . for all 75 The stage cost and all x ∈ X. closed. κf (x)) + (|w|). z). w). the following assumption is introduced. 2005].

4 is dierent from is that. where σ = pNp λ. 0)] w and its linearization at (4. 0)] w (4. z. w)| ≤ σ ε2 . z. z. z. there exists that σ( ¯ for all 2 w J(x. λ. and all w ∈ MW . ·) are C with respect to the variables w. w) T w J(x. note that C 2 is very dierent from requiring κM P C (x) is C 2 . control law that. u. Moreover. w)) ≤λ x ∈ X. could be discontinuous. ∞) such are compact. z.4 hold. 0)) ε . as discussed in Chapter 1.76 Chapter 4. ·). all u ∈ MU . 0) − [ J(x. under Assumption 4.2 x. ·) worth to note that Assumption 4. the requirement κ(·. and W λ ∈ [0. z. z. z) arg max [ = sign ( T w J(x. will be proposed. Note that Assumption 4. ·. Since sets a constant X . for example.1 Suppose that Assumptions 4. a priori is not guaranteed.4 Remark 4.1 and 4. It is 2 requiring V (x) is C . w) be C2 with respect to w. Let introduce now the following lemma. z. z and Let assume J(x. ·). and all w ∈ MW . Note that the results of this chapter rely only on the existence of the constant which means that the proposed MPC formulation can be im- plemented without requiring an estimation of its value.4 can be ensured. z. z.4. Let dene EL (x. w) J(x. the following assumption has to be introduced.4 Guaranteed bound of the max function In this section a new cost function that avoids to evaluate the maximum of FHCLG. 2 x ∈ X. for all Then |EL (x.8) w J(x. all u ∈ MU .7) as the error between the cost function w = 0. w) − J(x. if the 2 functions f (·. Min-Max NMPC: a min formulation with guaranteed robust stability 4. U . For this aim. Assumption 4. l(·. Lemma 4. Vf (·) and κ(·. Let denote wg (x.

4).1) with xt = x. t + Np − 1] w[t.3)-(4. z) as dened in (4. the terminal penalty Xf . The idea is to evaluate the cost function only for an appropriately calculated sequence of disturbances so reducing dramatically the computational burden of the maximization problem. the Finite Horizon Approximated Closed-Loop Game (FHACLG) problem consists in minimizing. for all k ∈ [t.2).5. the constraints (4. z. z) J(x.5 Proposed Min formulation: stability A new formulation of the min-max model predictive control is presented in this section. wg (x. Given ¯ Nc and Np . the state dynamics (4. 0) (remember J(x.9) σ = pNp λ. z) where (4. z) − σε2 ≤ J(x. which requires the solution of a maximization problem. k ∈ [t + Nc .2 Consider the FHCLG given in Denition 4. z) ≤ J(x. and the auxiliary control law κf .4. z.1 and 4. |w|∞ ≤ ε). that by constraint (4. t + Np − 1] .t+Nc −1] .2 (FHACLG) the positive integers Vf the terminal set Consider system (4. The new min-max problem Denition 4.1) 2. Under Assump- tions 4. 4. the uncertainty sequence to wg (x. Lemma 4. the stage cost l. the control signal uk = κ(xk . with respect z[t. ˜ J(x. zk ).5.4 ¯ ˜ ¯ J(x. Proposed Min formulation: stability 77 the uncertainty sequence maximizing the rst order approximation of the cost function Dene ˜ J(x. z). k ∈ [t. Consider the proposed cost function is stated in the following denition. Now it is possible to state the following result.t+Np −1] ∈ MW and all 3. The objective is to circumvent the computation of the worstcase disturbance realization. z)).8). z) subject to: 1. t + Nc − 1] κf (xk ). the cost function ˜ J(x.

10) u = κM P C (x) = κ(x. the optimal value of the performance index.1).3 Denote with V (x) and ˜ V (x) ˜ z J(x. ˜o ) the optimal solution of the FHCLG and the FHACLG problem respectively. zo be a solution of the FHACG. z0 ) ˜ ˜o where z0 ˜o is the rst element of ˜[t. a tube-based formulation for the robust constraint satisfaction could be used (see for example [Limon et al. the terminal state constraints xt+Np ∈ Xf . is employed as an ISpS-Lyapunov function for the closed-loop system (4. Min-Max NMPC: a min formulation with guaranteed robust stability 4. It is clear that. 2008]).e.10). Lemma 4. V (x) In what follows. ˜o ) ˜ ˜ V (x) ≤ V (x) ≤ V (x) + σε2 . the optimal solution ˜ zo of the FHACLG is a sub- optimal feasible solution for the FHCLG. for all w[t. Letting ˜ zo ˜[t.t+Nc −1] . zo Remark 4. 2. the computational burden of the maximization problem is dramatically reduced. the feedback control law u = κM P C (x) is obtained by setting ˜ (4.t+Np −1] ∈ MW . the constraints satisfaction must be checked for all possible disturbances. As it is claimed in the following lemma. according to the RH paradigm. (4.t+Nc −1] . since the constraints of the FHCLG and the FHACLG problems are the same. by solving the FHACLG instead of the FHCLG prob- lem. i. ˜o ) − σε2 ≤ V (x) ≤ J(x.1 and 4.78 Chapter 4. In the following. Under Assumption 4. the dierence between the optimal value of the original objective function and the value obtained with ˜ zo is bounded by σε2 . Assumption 4.4 it results that 1.5 Let .3 Even if. let ˜ X MP C denote the set of states for which a solution of the FHACLG problem exists. In order to guarantee this. ¯ z ¯ z J(x.

Theorem 4.1) is attached to the wall via a spring with elastic k given by k = k0 e−x1 .1-4. where x1 is the displacement w = 0. the proposed controller guarantees that the optimal original cost function is an ISpS Lyapunov function for the resulting closed-loop 4.13) (in o ¯ the denition replace d2 with w ) with function V (x) = J(x.1). 2003]).1 Under Assumptions 4. is contained in IΩ. z ). dened in (3. Example ˜ • Ξ = X MP C • Ω = Xf • α1 = αl • α2 = βVf • α3 = αl • λ2 = • c1 = Np (ε) • c2 = σε2 79 ε). This carriage (see Figure 4. is ISpS with respect to ˜M P C .4). Finally of the carriage from the equilibrium position associated with the external force u=0 and the external disturbance force (wind force) .5.4 system.6 Example In this section. The control applied at time t is κM P C (x) ˜ and not κM P C (x). the closed-loop system formed by (4. subject to constraints (4.2)-(4. The set is such that the set W Θ (depending from The main result can now be stated. In spite of this. with RPIA set X w Remark 4. the MPC law introduced in the chapter is applied to a cart with a mass M moving on a plane (the model is the same of the paper [Magni et al.10).6. (4.4.

has to choose the value of the parameters z1 and z2 . ueq = 0. as described in Section 4. with Q and R the same used for the LQR. The problem of solving the maximization of the FHCLG in exact manner is unrealistic. The wind force is assumed to be bounded. The a damper with damping factor model of the system is given by the following continuous-time state space nonlinear model x1 (t) = x2 (t) ˙ k0 x2 (t) = − M e−x1 (t) x1 (t) − ˙ where hd M x2 (t) + u(t) M + w(t) M x2 is the carriage velocity. at each instant time. The K = [0.7159]. so that only an approximated solution is possible in real-time. The pro- posed strategy.4. An Euler approximation of the system with sampling time Tc = 0.4 s is given by x1k+1 x2k+1 = x1k + Tc x2k k0 = −Tc M e−x1k x1k + x2k − Tc hd x2k + Tc uk + Tc wk M M M which is a discrete-time nonlinear system. The parameters of the system are M = 1 kg .33 N .65 m.4 N . The stage cost adopted in the cost function is quadratic. obtained for the system linearized in xeq = 0.9091 2.1} with (s) = 10s.2. con- . However. at each time. hd = 1. The feedback control strategies are assumed to have the following structure k(x. Min-Max NMPC: a min formulation with guaranteed robust stability hd aects the system in a resistive way. and denote with P the unique symmetric positive denite solution ˜ of the Lyapunov equation (A − BK) P (A − BK) − P + Q = 0 auxiliary control law with κf = −Kx.8639 3. For this system the MPC control law is computed according to the algorithm presented in the paper.3 is satised in the terminal set Xf = {x : x P x ≤ 6.80 Chapter 4. with identity matrix of dimension 2 × 2 and R = 1 Let Q β = 1. is the LQR 2×2 . The lengths of the control and prediction horizon in the MPC implementation are respectively Nc = 6 and Np = 10. k0 = 0.6616 Vf (x) = x P x. the Assumption 4.5 N . This means that the minimization problem. x Qx + u Ru. m m |w| ≤ 0. using Q = I ˜ β(Q + K RK). |x1 | ≤ 2. nds the maximum of the linear approximation of the cost function.5332 0. P = Using as terminal penalty 5.1 N s . z) = −z1 κf (x) + z2 . The state and control variables have to satisfy the following constraints |u| ≤ 4.8639 2.

. The obtained results show the stability and admissibility of the proposed controller. A tube-based formulation for the robust constraint satisfaction could be used to check that. 4.1: Cart and spring example. the solution presented in the paper requires only one evaluation. constraints are never violated.61m. of all the vertexes of the disturbance. which requires the solution of a maximization problem. which in this case are 2Np = 210 = 1024. s with a particular realization of the disturbance. this has been tested on a model. The proposed controller is based on the simple evaluation of the cost function for an appropriately calculated sequence of future disturbances. The implementation of the MPC for the illustrative example demonstrates the diculty to implement a real min-max MPC compared with the proposed one.2 shows the evolution of the system for the initial state x1 = −2. The controller derived from the minimization of the relaxed maximization problem ensures the same domain of attraction of the min-max one and guarantees ISpS under the same conditions. Conclusions 81 sidering the case of a cost function convex in w. at each time instant. The dashed lines show the minimum and maximum values corresponding to all possible disturbances along the evolution of the system. the exact solution of the max requires the evaluation.7 Conclusions A new formulation of the min-max model predictive control has been presented in this chapter. Its objective is to make aordable the computation of the worst-case disturbance realization. x2 = 0 m . Instead. for all possible sequence of disturbances.7. The following Figure 4. In order to illustrate the proposed controller.4. Figure 4.

w))w w. z.2 2 1 0 −1 −2 −0. 0)] w 1 + 2 wT [ 2 J(x. w)]w 2 T ˆ w J(x.4 0. . w) = wT [ 2 1 ¯ |EL (x.8 Appendix Proof of Lemma 4. w) = J(x. z. z. z. Min-Max NMPC: a min formulation with guaranteed robust stability 5 2 4 1 position velocity 0 2 4 6 8 time 10 12 14 16 3 0 2 −1 1 −2 0 −3 −1 0 2 4 6 8 time 10 12 14 16 (a) Position of the cart (b) Velocity of the cart 5 4 3 0. z.1 0 −0.2 −3 −4 −5 0 2 4 6 8 time 10 12 14 16 −0.1 0 2 4 6 8 time 10 12 14 16 (c) Control applied to the cart (d) Wind force Figure 4.82 3 Chapter 4. 0) + [ This means that T w J(x. w ∈ MW there exists a w ∈ MW ˆ such that J(x.1: for every using the Maclaurin series of the function J(x. w)]w. z.3 −0.61m.3 0. w)| ≤ σ ( 2 2 ˆ w J(x. x2 = 0 m s 4. z. ˆ w 1 EL (x.2: Evolution of the system for initial state x1 = −2. w).4 Noise Input 0. z. z.

z. 0)] w. z) w∈MW . w) = EL (x. z. z)) + σε2 . maximizing the cost with respect to w.7) EL (x. w) + J(x. z))| ≤ σ ε2 + σ ε2 = σε2 . Appendix σ( ¯ and 83 2 V w Assumptions 4. equation (4. w) = EL (x. z. z)) = J(x. wg (x. w) ≤ J(x.4 guarantee that all (x. z.4. z. w) w∈MW ≤ max J(x. z. z. w)| ≤ λpNp λpNp 2 σ 2 λ T w w≤ |w|2 ≤ ε = ε . one has ¯ J(x. ˆ w ∈ IRp w ∈ IRpNp . 0)] w + EL (x. z)) + σε2 = J(x. z. z. z. z. z. z)) ≤ |EL (x. z. wg (x. 0)]T w +EL (x. z. In order to conclude the proof. ˆ one has u ∈ MU and all w ∈ MW . z. wg (x. wg (x. z) − σε2 ≤ J(x. w) ≥ J(x. Using Lemma 4. for all x ∈ X . z. wg (x. z). wg (x. z) ≤ J(x. z. z) + EL (x. and considering (4. z.1 and 4. z. 0) + [ ˜ = J(x. one obtains EL (x. note that (4. 0) + [ w∈MW T w J(x. z. z. z)). z) = max J(x. ∞ 2 2 2 2 Proof of Lemma 4.1. wg (x.2: by denition. z) + σε2 and hence T w J(x. 0) + [ w J(x. z)) J(x. ¯ ˜ J(x. z. Now. z)) + σε2 ¯ ˜ J(x. 0)] w + EL (x.11) since ˜ ¯ J(x. w)| + |EL (x. w) + J(x.8). wg (x. z. w) − EL (x. z)) − EL (x. remembering that |EL (x. z) = max J(x. z. 2 2 Then J(x. J(x. z. z. wg (x. z. By this. z. 0)] wg (x. 0) + [ Adding and subtracting the term T w J(x. z. z. 0) + [ T w J(x.8. z). w)) ≤ λ.

zo ) ≥ J(x. The second inequality stems directly from the suboptimality of ˜ zo : V (x) ≤ ¯ ˜ J(x. zo ) = V (x) and the second one ˜ ˜ ˜ ¯ ˜ ¯ V (x) = J(x. . k ∈ [t + 1. Np ) at time t.9) the rst inequality ¯ ˜ ˜ ˜ ˜ V (x) = J(x. when not strictly necessary.9) ¯ ˜ ˜ ˜ ¯ ˜ V (x) = J(x.2.3 First claim: denoting zo the optimal solution of the FHCLG and ˜ zo the optimal solution of the FHACLG. zo . equation (4. the arguments Nc and Np will be omitted. t + Nc − 1] zf .t+Nc −1] zo is a RPIA set for the closed-loop system. Np ) = In the following it will be shown that the function ¯ J(x.2. Nc .3 by taking zs = zk . the rst inequality is obtained by Lemma 4. applying κf (x) at time t+Np . ˜ X MP C. zo ) ≥ J(x.1).3. zo ). of the FHACLG at time t. t + Np ] t + 1 starting from the sequence ˜ z. k ∈ [t + Nc . the policy parameters zs is admissible. Min-Max NMPC: a min formulation with guaranteed robust stability Proof of Lemma 4. For brevity. there is xt+Np ∈ Xf . zo ) ≥ J(x. Np ) is easily derived from AssumpThe robust invariance of X tion 4. zo ) − σε2 = V (x) − σε2 and hence ˜ V (x) ≤ V (x) + σε2 . xt+Np +1 ∈ Xf . Second claim: again. for all w ∈ MW ˜ M P C (Nc . V (x. zo ) − σε2 ≥ J(x. Nc . if system admits an ISpSthen it is ISpS in Proof of Theorem 4. therefore X as policy parameter vector at time ˜[t. Np ). Nc . (4. zo ) ≥ J(x. By Assumption 4. for all w ∈ MW .84 Chapter 4.1: Lyapunov function in ˜ X MP C. zo ) ≥ J(x. (4. from (4. ˜ M P C (Nc . By the optimal solution = arg minz J(x. zo ) − σε2 . Np ) and hence. zo ) ≥ J(x.10) in is an ISpS-Lyapunov function for the closed-loop system ˜ X M P C (Nc . Np ). by Theorem 3.

w[t. Np ) ≤ V (x.4. Np ) + (ε) which holds for all ˜ x ∈ X M P C (Nc . ˆ J(x. z0 )) ≥ αl (|x|) for all (4. Nc + 1. Appendix 85 Then. Np + 1) ≤ t+Np −1 l(xk . and all w ∈ MW . z. Nc + 1. uk ) + Vf (xt+Np ) − Vf (xt+Np ) + Vf (xt+Np +1 ) k=t +l(xt+Np . κ(x. Np ).3. ≤ V (x. using Assumption 4. Nc + 1.8.t+Nc ] . . z[t. Np + 1) ≤ ≤ w∈MW ˆ max J(x.13) x ∈ Xf . Nc .t+Np ] . Np − 1) + (ε) ≤ . uk ) + Vf (xt+Np ) + (|w|) k=t which implies V (x. zo . uk ) + Vf (xt+Np ) + (|w|) k=t = V (x.3 ˆ J(x. . Nc + 1. 0) z ≥ l(x.t+Nc ] . z[t. Nc − 1. Np + 1) = t+Np −1 l(xk .2. policy vector for the FHCLG at Then In order to derive the upper bound. Nc . w[t. z[t. Therefore. the lower bound is easily obtained V (x) = J(x. wo ) ≥ min J(x. consider the following policy vector ˆ z[t.12) ˜ x ∈ X MP C.t+Nc −1] zf ] as admissible time t with horizons Nc + 1 and Np + 1.t+Nc ] [zo [t. 0) + Np (ε) = Vf (x) + Np (ε) ≤ βVf (|x|) + Np (ε) for all (4.t+Nc ] . w[t.t+Np ] . the upper bound is obtained V (x.t+Np ] . κf (xt+Np )). . using Assumption 4. Np + 1) t+Np −1 w∈MW max l(xk . In view of Assumption 4.

zo . it is possible to derive that V (xt+1 ) − V (xt ) ≤ −l(xt . Np ) to and hence.10) in ˜ X M P C (Nc . by Assumption 4. κM P C (xt )) + (|wt |). zo .14).3 ˜ ˜ ¯ J(xt+1 . w) − J(xt . zo . w) ≤ −l(xt . by Theorem 3. zs . by (4. in order to nd the bound on w = arg max J(xt+1 . .5.4). ¯ ˜ ˜ From the feasibility of ˜ . w) ≤ Vf (xt+Np +1 ) − Vf (xt+Np ) +l(xt+Np . κM P C (xt )) + (|wt |) + σε2 . Min-Max NMPC: a min formulation with guaranteed robust stability ∆V . wt+1 . wt+Np −2 ). · · · . in view also of Assumption 4.1). w) ≤ J(xt . Np ) ˜ w with zs dened as previously.3 and by the optimality of ˜ zo ¯ ˜ ˜ ¯ ˜ ˜ J(xt . w. and all w ∈ MW .2. Np ). Nc .2)-(4. V (x) Therefore.10). κM P C (xt )).13). = V Therefore. In virtue of Lemma 4.12). and take w = (wt . zs . subject to constraints (4.1) and (4. (4. w) = J(xt+1 .86 Chapter 4. zs ) ≥ V (xt+1 ).14) ˜ x ∈ X MP C. ˜ Then. dene Now. is ISpS with respect w with RPIA set ˜ X M P C (Nc . κf (xt+Np )) −l(xt . (4. ˜ for all (4. zo ) + σε2 ˜ (xt ) + σε2 ≤ V (xt ) + σε2 . it is easy to derive that zo ˜ ˜ ¯ J(xt+1 . zo ) ≤ J(xt . is an ISpS-Lyapunov function for the closed-loop system (4. ˜ By optimality of w ˜ ¯ ˜ J(xt+1 . zs . (4. w) − J(xt . the closed-loop system formed by (4. zs .

. .3. A . . . 87 89 92 Problem formulation Robust MPC strategy . . .1 5. . . . . . . . . . . . . . . . . . it is necessary to introduce Min- some knowledge on the uncertainty in the optimization problem. . .1 Introduction Nominal MPC. . . . . . . . . . .3 5. . . 2003. . . . . . . . .2 5. . . . . . . max model predictive control. Appendix . . . . . . . . . . . . . Conclusions . Magni et al. . even if it has important inherent robustness properties in the unconstrained case. . . . . . . . .Chapter 5 Robust MPC of Nonlinear Systems with Bounded and State-Dependent Uncertainties Contents 5. . . . . Scokaert & Mayne 1998. . . . .3. . . . . . . . . . . . Regional Input-to-State Stability . . . . 5. . In order to achieve this goal. . 93 94 95 96 99 99 Feasibility . . . . . . . Bemporad et al. . . . . cannot guarantee the robust satisfaction of state constraints. . . . . Simulation Results .1 5. .6 Shrunk State Constraints . . .5 5. . . .4 5. . . . .3 Introduction . . . . . . . . .2 5. . . .3. 2003]). . . . . . . . . . . . . . . . . . . . . . . . . . 5. . . . . . is very computational demanding and the range of processes to which it can be applied is limited to those that are small-size and with very slow dynamics (see for example [Lee & Yu 1997. . . . . . . . . .

thus accounting for a possible reduction of the state dependent component of the uncertainty due to the control action. and the state is limited in a compact set. On the contrary. Robust Open-loop NMPC more ecient technique is based on solving the nominal open-loop optimization problem using tightened state constraints in order to guarantee that the original constraints are fullled from the real system for any possible realization of the uncertainties. If the system is aected by state dependent disturbances. All this works consider additive distur- bance. The goal of this chapter is to modify the algorithm presented in [Limon et al. Rakovi¢ et al. along which the propagation of the uncertainty must be taken into account. 1) the horizon. in order to limit the spread of the trajectories. However. it is always possible to nd the maximum value of the disturbance and to apply the algorithms described in [Limon et al. Raimondo & Magni 2006. In this regard. The idea was introduced in [Chisci et al. The restricted sets are computed on-line iteratively by exploiting the state sequence obtained by the open-loop optimization. the concept of regional ISS is used in order to show that the obtained closed-loop system is regional ISS with respect to the bounded persistent disturbance. a long prediction horizon. a closed-loop term was considered for nonlinear systems in [Rakovi¢ et al. is useful to better approximate the performance of the Innite Horizon control law (see e.g. signicative advantages can be clearly obtained. The main drawback of this open-loop strategy is the large spread of trajectories along the optimization horizon due to the eect of the disturbances. 2002a] and [Raimondo & Magni 2006]. 2006b].88 Chapter 5. 2001] for linear systems and applied to nonlinear systems in [Limon et al. In order to reduce this. Moreover. 2001a]). where the concept of tube is explored. [Magni et al. it is possible to show that the devised technique yields to an enlarged feasible region compared to the one obtainable if just an additive disturbance approximation is considered. 2002a] in order to eciently consider state dependent disturbances. if the particular structure of the disturbance is considered. 2) the terminal constraint is imposed at the end of the control horizon. is reduced by using a control horizon shorter than the prediction one. The robustness with respect to state dependent disturbance is analyzed using the stability margin concept. 2006b]. . 2002a. In order to analyze the stability properties in the presence of bounded persistent disturbances. in the proposed algorithm.

υk ). we will not point out the functional dependence of its arguments except where strictly needed.e. 0) = 0.5. uk ∈ IRm the control vector r and υk ∈ Υ ⊆ IR an exogenous input which models the disturbance. k ≥ 0 .2 Problem formulation Assume that the plant to be controlled is described by discrete-time nonlinear dynamic model xk+1 = f (xk . For simplicity of notation. uk ) ∈ IRn discrete-time state transition uncertainty. The state and control variables are restricted to fulll the following constraints x∈X u∈U where (5. there exists a positive constant Lfx such that (5. i.1 1.5) ˆ ˆ |f (a. where (5. 3.1) xk ∈ IRn denotes the system state. Given the system (5. such that ˆ xk+1 = f (xk . The map ˆ f : IRn × IRm → IRn is Lipschitz in x in the domain X × U.2.e. uk . b ∈ X and all u ∈ U. u) − f (b. uk ). denote the nominal model used for control design purposes.3) (5. υk ) ˆ f (xk . with ˆ Υ compact set containing the origin. it is assumed that the origin is an equilibrium point. In the sequel. ˆ with f (0. dk (xk . 0) = 0. u)| ≤ Lfx |a − b| for all a. ˆ f (0. for the sake of brevity. respectively. k ≥ 0 .2) denotes the dk = dk (xk . 2. υk ) − f (xk . Assumption 5. i.1). con- taining the origin as an interior point. where (5.4) X and U are compact subsets of IRn and IRm . . uk . υk ) on The system is supposed to fulll the following assumptions. uk . Problem formulation 89 5. uk ) + dk . uk . let f (xk .

uk . that is dk (xk .k+j−1|k] . with Finite Horizon Optimal Control Problem (FHOCP) consists in minimizing. with respect to u[t. rst let u[t2 . the terminal set XNc and a sequence of conˆ sets Xk|t ⊆ X . functions.6) ˆx xk+1 = f (ˆk . . let xk+j|k ˆ ˆ(xk . The K-function δ is such that Lδ L|x|. an straint Consider system (5.1) with xt = x. Np .t+Nc −1|t] . uk ). uk ). Assumption 5. t + Nc − 1] (to be described later on) the Nc . The control objective consists in designing a state-feedback control law capable to achieve ISS closed-loop stability and to satisfy state and control constraints in presence of state dependent uncertainties and persistent disturbances.t+Nc −1|t] . . . Denition 5.1 (FHOCP) positive integer Vf . .2 The additive transition uncertainty dk is limited in a time varying compact ball Dk . and all υk ∈ Υ. be the predicted state at k + j obtained with the nominal model f with initial condition xk and input u[k. ut3 |t1 ]. . ∀x∈X } exists nite. υk ) ∈ Dk for all B(δ(|xk |) + µ(Υsup )) where uk ∈ U . uk|t ) + x l(ˆk|t . Then. be a control sequence. u[t. In order to introduce the MPC algorithm formulated according to an [ut2 |t1 ut2 +1|t1 . the performance index J (¯.t3 |t1 ] Horizon Optimal Control Problem (FHOCP) can be stated. given k ≥ 0. Moreover. the terminal penalty auxiliary control law κf . δ and µ are two Kmin{L ∈ IR>0 : δ(|x|) ≤ bounded by the sum of two contributions: a state-dependent component and a non-state dependent one. with t1 ≤ t2 ≤ t3 . ˆ with xt = x ˆ ¯ . Given the ¯ Nc ≤ Np . the stage cost l. The state of the plant can be measured at each sample time. k ∈ [t. κf (xk|t )) + Vf (ˆt+Np |t ) x x (5. It follows that dk is all xk ∈ X . j ≥ 1. Nc . Robust Open-loop NMPC xk 4. the following Finiteopen-loop strategy. Np ) x t+Nc −1 k=t subject to 1. .90 Chapter 5. the nominal state dynamics t+Np −1 k=t+Nc l(ˆk|t .

for all x ∈ X .2. . the control and prediction horizon Nc and Np . it will be shown how to choose accurately the stage cost l. the state constraints ˆ xt+j|t ∈ Xk|t . dene κM P C (¯) x where uo (¯) t|t x and apply the control law uo (¯) t|t x is the rst column of uo [t. u = κM P C (x).3 t. . note that.t+Nc −1|t] by solving the FHOCP. . . t + Nc − 1]. . Then. Problem formulation k ∈ [0. The stage cost denes the performance index to optimize and satises the following assumption. x for all k ∈ [t + Nc . . the control constraint (5. all u ∈ U . ¯ nd the optimal control sequence uo [t. . the terminal constraint XNc and the auxiliary control law κf in order to guarantee closed-loop ISS. Another peculiarity is the use of a state constraint that changes along the horizon. where l is a K∞ -function. l(x. u) is such that l(0. l(x. in this case the constraint sets are dened only within the control horizon and the terminal constraint is stated at the end of the control horizon. Moreover. . ˆ k ∈ [t. the constraint sets ˆ Xk|t . 0) = 0. . given xt = x. the auxiliary control law xt+Nc |t ∈ XNc ˆ uk|t = κf (ˆk|t ).7) With particular reference to the underlined denition of the FHOCP. u). . the terminal state constraints 5. . (5.4) for all 3. In particular the set XNc will be chosen such that. u) is Lipschitz with respect to x and u in X ×U . The stage cost function The usual RH control technique can now be stated as follows: at every time instants Assumption 5. Np − 1] for all 91 2. according to the Receding Horizon (RH) strategy.t+Nc −1|t] . . the terminal cost function Vf . . . . t + Nc − 1] 4. In the following. . t + Np − 1]. k ∈ [t.5. with Lipschitz constants Ll ∈ IR≥0 and Llu ∈ IR≥0 respectively. and l(|x|) ≤ l(x. starting from any x ∈ XNc in Np − Nc steps the auxiliary control law can steer the state of the nominal system into a set Xf which satises the assumption asked for the terminal set of standard . with respect to the usual formulation.

7. for all x ∈ XNc . for all [κf (ˆt|t ). κf (x)) ∈ Xf . X MP C will denote the set containing all the state vectors for which a fea- sible control sequence exists. the closed loop map ˆ f (x.e. 2000]. Assumption 5. and a set Xf . Xf κf (x) ∈ U . i. there exists a positive scalar κf ε ∈ IR>0 such that ˆ f (x.92 Chapter 5. x x x ˆ xt ∈Xf . a control sequence all the constraints of the FHOCP. κf (x)).4 design parameters 1. ˆ f (x. . 2. with xt|t =xt . . κf (x) is Lipschitz in Xf . ˜ u[t. Lκf ∈ IR>0 IR>0 closed. is Lipschitz in Xf x ∈ Xf with constant with constant Lfc ∈ 4. let introduce the following further assumptions. for all Xf in Np − Nc steps under the nominal dynamics in closed-loop with the auxiliary control law 3. . 5. LVf ∈ IR>0 x∈Xf ˆ Vf (f (x. . Robust Open-loop NMPC stabilizing MPC control algorithm [Mayne et al. 6. κf (x)))−Vf (x)≤−l(x. κf (x)) ∈ XNc B(ε). Assumption 5. with constant 3. u[t. Given an auxiliary control law κf . the Vf is such that Xf ⊆ X . The robust terminal constraint set of the FHOCP. Vf (x) for all is Lipschitz in Xf . XNc ⊇ Xf x ∈ XNc the state can be steered to 2. κf (ˆt+1|t ). for all 0 ∈ Xf x ∈ Xf .5 is chosen such that 1. κf (ˆt+Np −1|t )] . In the following.t+Np −1|t] for all is a feasible control sequence for the FHOCP. XNc .3 Robust MPC strategy In order to formulate the robust MPC algorithm.t+Nc −1|t] satisfying 5. κf (x)).

3.5. ¯ t+1|t t+N trol action implicitly ¯ u[t+1. Lemma 5. . . . with xt|t =xt .8) ρt+1|t ˆ µ + Lδ |xt |. . u◦ c −1|t . with Then. under Assumption 5. ¯ u[t. computed as follows ˆ Xk|t where X B(ˆk|t ).3. . x x x ˆ xt ∈ XNc . Remark 5.1 Shrunk State Constraints In the following. given the state vector xt at t. let a control sequence.1). . for any admissible disturbance sequence.9) ¯ u[t. the sequence µ ¯ µ(Υsup ). u]. t + Nc − 1] ρ ¯ x (5.2. κf (ˆt+Np −1|t )] . . for all k ∈ [t + 1.t+N c −1|t] . compared to previous apIn fact. . t + Nc − 1]. k ∈ [t + 1. for all [κf (ˆt|t ). . it relies on xk|t . . by imposing restricted constraints to the predicted open-loop trajectories. 5. .8).1 and 5. the following notation will be used: given ◦ an optimal sequence u [t. given state prediction error will be derived. and all υ ∈ MΥ .t+Nc |t+1] u ∈ U is a suitably dened feasible con¯ depending on xt+Nc |t+1 . . .1 formation The constraint tightening (5. .2. xt at time t. . ˜ u[t. ¯ ρk|t ˆ = (Lδ + Lfx )ˆk−1|t + µ + Lδ |ˆk−1|t |. j ∈ [t + 2. κf (ˆt+1|t ). all xt ∈ X M P C . . it is shown that the satisfaction of the original state constraints is ensured. .t+Nc −1|t] of control actions obtained by solving the FHOCP at time [u◦ . The . . .1 time Under Assumptions 5. Raimondo & Magni 2006]. The following result will be ˆ let dene the sequence where t. rather then using only the state inthe whole predicted state sequence accounting for a possible reduction proaches [Limon et al. applied to the perturbed c system (5. be feasible with respect to the restricted state constraints of the FHOCP.t+Np −1|t] is a feasible control sequence for the FHOCP. a norm-bound on the Subsequently. guarantees xk ∈ X . 2002a. thus ˆ of the state-dependent component of the uncertainty along the horizon. ρ (5. t + Nc − 1]. xt .t+N −1|t] . Throughout this section. ˆ Xk|t . may lead to less conservative computations. Robust MPC strategy 93 4. . instrumental for the subsequent analysis.

XNc ). 2006a]).e.2 Feasibility In order to show that X MP C is a RPIA set. Given a set n subject to (5. 2. 2005. 2002.4). Lemma 5. an upper norm bound for the maximal admissible uncertainty will be stated in Assumption 5.6. is a RPIA set under the Now it will be stated and proved that closed-loop dynamics. X MP C. Given a set and a positive constant ε as in Assumption 5. u) ∈ Ξ . let dene ¯ d dist(IRn \P(XNc ). Kerrigan et al. ut ). In this re- gard.3. fx Assumption 5. It must be remarked that.2. 5. Rakovi¢ et al. P(Ξ) is ˆ the set of states which can be steered to Ξ by a control action under f (xt . motivated by the following Denition 5. Robust Open-loop NMPC eectiveness of the proposed approach in enlarging the feasible region of the FHOCP will be shown in Section 5. Then.4 by a simulation example.94 Chapter 5.1 and 5. ¯ L1−Nc d. ˆ P(Ξ).5.. is dened as P(Ξ) x ∈ IR ∃u ∈ U : f (x. ¯ dκf XNc and ε/Lfx ¯ XNc ⊂ XNc ⊕ B(dκf ) ⊆ P(XNc ). the (one-step) predecessor set.2 Consider that Assumptions 5. although the underlying theory is well established and many dierent methods have been proposed since the seminal paper [Bertsekas & Rhodes 1971].7).5 hold.1) and (5. there exist ecient numerical procedures for the computation of pre-images and predecessor sets (see [Bravo et al. it holds that 1.6 for all The K-functions δ and µ are such that δ(|xt |)+µ(Υsup ) ≤ xt ∈ X . ¯ ¯ d ≥ dκf . Denition 5. under the closed loop dynamics given by (5. for some classes of nonlinear systems. for nonlinear systems. .2 (P(Ξ)) Ξ ⊂ X .2 and Lemma 5. i. the numerical computation of P(XNc ) is a very dicult task.

8 l(|xt |) − ϕx (|xt |) LNc −1 x The stage transition cost is a K∞ -function for all l(x.1) and (5.3.t+Nc −1|t] . fx Assumption 5. Robust MPC strategy Theorem 5. The solution of closed-loop system formed by (5.7) and x=0 ¯ w=0 with respect to disturbances and initial In what follows.6. 2002a. the possibility to compute ¯ d relying on P(XNc ) XNc Xf as stabilizing constraint set. (5.2).7) are analyzed. Raimondo & Magni 2006].7).3 Regional Input-to-State Stability In the following. the use of them along a longer horizon (the entire prediction horizon) could reduce the feasible set of the FHOCP. u). Under Assumptions 5.3. allow to en- large the bound on admissible uncertainties which the controller can cope with. Nc . fx ¯ L1−Nc dκf .7 is continuous at conditions.1-5.10) is employed as an ISS-Lyapunov function for the closed-loop system formed by (5. Assumption 5. the set in which the FHOCP M P C . are such that δ(|xt |)+µ(|υt |)≤ . uo [t. considering that the restricted constraints are based on Lipschitz constants that are conservative since globals (or just regionals due to the presence of constraints). Remark 5.3) and (5. In fact.1) and subject to (5.2 and the use of With respect instead of to previous literature [Limon et al.5.e. with ϕx (|xt |) Lfcp N −(Nc +1) Ll Lfx −1 + LVf Lfcp f N −(Nc +1) LNc + (Ll + Llu Lκf ) fx −1 Lfc −1 LNc δ(|xt |). X of the control law given by (5. the optimal value of the performance index. Assumption 5.1 95 Let a system be described by equation (5.4). 5. is also RPI for the closed-loop system under the action is feasible.7). i. Np ) (5. V (x) J(x. the stability properties of system (5.1) in closed-loop with (5.9 for all The K-functions δ and µ xt ∈ X and all υt ∈ Υ. is such that α3 (|xt |) xt ∈ X M P C .

Under Assumptions 5. .8 N −N −1 • σ(s) = Ll Lfx −1 + (Ll + Llu Lκf ) f The set c Lfcp Lfc −1 LNc + LVf Lfcp fx N −(Nc +1) LNc µ(s) fx Υ is such that the set Θ (depending from with function V given by (5.5 (1+ x1k ) uk ] + d1k x2k+1 = x2k +0. for implementation purposes the balls to be subtracted (in the Potryagin sense) from the constraint set X to obtain ˆ Xk|t . . .11) is subject to state and input constraints (5. LNc −1 x Chapter 5.1. System (5.96 Assumption 5. The uncer- tainty vector is given by dk =1·10−3 xk +υk .7) is regional ISS in X Theorem 5.3) and (5. . . 5.2 v ∈ Υ.1-5.10.05 [x1k + 0.4 Simulation Results Consider the following discrete-time model of an undamped nonlinear oscillator x1k+1 = x1k +0.17).05 [−x2k + 0. for all k ∈ [t + 1. IΩ. Robust Open-loop NMPC Let with ϕx (s) as in Assumption 5. while stant of the system is X The Lipschitz con- Lfx =1.10 • Ξ = X MP C • Ω = Xf • α1 (s) = l(s) • α2 (s) = LVf s • α3 (s) = l(s) − ϕx (s).11) (i) denotes the i-th component of a vector.5 (1− 4x2k ) uk ] + d2k where the subscript (5. Since aordable algorithms exist for the numerical computation of the Pontryagin dierence set of polytopes.1) under the acM P C with respect to tion of the MPC control law (5. the system (5. U {u ∈ IR : |u| ≤ 2}. where the set is depicted in Figure 5.1390. t + Nc − 1] are outer approximated by convex parallelotopes.10). with |υ| ≤ 1·10−4 . is contained in Υsup ).4). dened in (2.

12 x≤1 x∈IRn : x −43. R = 1. u) and the nal cost Vf by Vf (x) −23.5955 0. respectively k = [0.21 −29. can be designed as described in [Parisini et al.45 x≤1 x∈IRn : x −29. Nc =8.50 XNc 114. Choosing the following ellipsoidal sets.9764] and Xf and XNc .11) in a neighborhood of the origin.0479 · 10−4 .67 and with by Lκf = 1. 1998]. . Figure 5. A linear state feedback control law with stabilizing (5. Simulation Results u = κf (x) = k x.28 x P x.61 then Assumption 5. satisfy Assumption 5.4 is satised.45 208. .56 −23.4 and 5. are bounded by X MP C δ(|xt |) + µ(Υsup ) ≤ 5.1437.21 −43. Q = 0.5 Xf 167. 97 k ∈ IR2 . for which the feasibility set the closed-loop dynamics.3311 · 10−4 ¯ d=1.61 167.1 0 0 0.5. . be computed: The following values for and ¯ dκf and ¯ d can ¯ dκf =2. for all xt ∈ X . P = 91.1 Np =50.12 305. It follows that the is RPI under admissible uncertainties. Lfc = 1.2554 · 10−3 . Let the stage cost l be given with l(x.2 shows the closed-loop trajectories of the system under nominal conditions (dashed) and with uncertainties (solid).0504 x Qx + u Ru.4.

2 0. −0.3 0.98 Chapter 5.1 Figure 5.04) .1 XN c −0.08 (c) 0. −0. (c)=(0.04 (a) −0.1 (a) x(2 ) 0 Xf XN c −0.04 x(2 ) 0 Xf (b) −0. Robust Open-loop NMPC 0.4 Figure 5.11.1: Three examples of closed-loop trajectories with initial points: (a)=(−0.1 (c) −0. 0.2 −0.2 X 0.05 0 x(1 ) 0.2: Confrontation of closed-loop trajectories without model uncertainty (solid) and with state-dependent uncertainty (dashed) . (b)=(0. the robust constraint set XNc and the set Xf 0.2) .2.08 −0.2 −0.1 0.06) .05 0.1 0 (b) x(1 ) 0. straint set T T T The state conare emphasized X.

the closed-loop system under the action of the MPC control law is shown to be ISS under the considered class of uncertainties. 5.2. .12) with (5. a nonlinear stability margin with respect to state dependent uncertainties is given. it follows that turn proves the statement. ˆ ˆ with k∈[t + 1.1).2: Notice that. Conclusions 99 5. e ˆ which in Proof of Lemma 5.9).5. t + Nc − 1].1: the rst Given xt .12) |ˆk|t | ≤ ρk|t . employing a constraint tightening technique. comparing (5.5. given a vector ¯ x ∈ XNc ⊕ B(dκf ). by Assumptions 5. . e ¯ x Finally. . is upper bounded by ˆ ˆx |ˆk|t | = |f (xk−1 . Future research efforts will be devoted to further increase the degree of robustness of the MPC control law.1 and 5. Then. Finally. to enlarge the class of uncertainties. (5. the robust con- straints satisfaction is guaranteed for the considered class of uncertainties. a robust MPC controller for constrained discrete-time nonlinear systems with state-dependent uncertainty and persistent disturbance is presented. the prediction error ek|t xk − xk|t . .6 Appendix Proof of Lemma 5.5 Conclusions In this chapter. In particular. . uk−1|t ) + dk−1 − f (ˆk−1|t . uk−1|t )| e ≤ Lfx |ˆk−1|t | + |dk−1 | ≤ Lfx |ˆk−1|t | + µ + Lδ |xt+j−1 | e e ¯ ≤ (Lfx + Lδ )|ˆk−1|t | + µ + Lδ |ˆk−1|t |. consider the state xk obtained applying k−t−1 elements of a feasible control sequence ¯ u[t. Furthermore.t+Nc −1|t] to the uncertain system (5. under suitable assumptions. to allow for less conservative results and nally to address the unavoidable approximation errors involved in the computation of the optimal control actions.

t+Nc |t+1] = xt+Nc +1|t+1 ∈ XNc . Then xt+Nc |t+1 ∈ P(XNc ).5.1.3 xt+1 = and and and Suppose that Assumptions 5. .2. fx f c c f Proof: Given 5. N ¯ |ξ| ≤ |ˆt+Nc |t+1 − xt+Nc |t | ≤ Lfxc −1 (δ(|xt |) + µ). t+1|t t+N and initialized with Moreover.2 and Assumption 5.6 hold. Since ξ ∈ B LNc −1 (δ(|xt |) + µ) ¯ fx xt+Nc |t ∈ XNc . κM P C (xt )) + dt . if xt+Nc |t + ξ = xt+Nc |t+1 ∈ P(XNc ). then. a possible feasible control sequence is given by where ¯ u[t+1. Using also Lemma 5.2. . ˆ ˆ ¯ ˆ ¯ µ(Υsup )≤L1−Nc dκ . ˆ δ(|xt |) + Proof of Theorem 5. u◦ c −1|t ]. u◦ c −1|t . by Assumption 5. and hence x ∈ c P(XNc ). xt+N |t+1 ∈ XN ⊕ B(dκ ). for all xt exists a feasible solution of the FHOCP at time instant the optimal solution in ∈ X M P C . t. there exists at least one vector such that Since ˆ f (x . Let introduce the following lemma that will be used in the proof of Theorem 5. xt ∈ X M P C .1: It will be shown that the region X MP C is a RPI set for the closed-loop system. the proof will be satisfy the robust constraint divided in two steps.t+Nc −1|t] xt|t = xt ˆ δ(|xt |) + ¯ u[t+1. Given xt and ˆ f (xt . . by Assumptions ˆ ˆ Hence. ˆ u◦ [t. In particular. κf (x )) ∈ XNc B(ε). if xt+1|t+1 = xt+1 . ˆ obtained respectively using the input sequences u◦ [t. xt+Nc |t+1 ∈ XNc ⊕ B(dκf ). . consider the predictions xt+Nc |t ˆ xt+Nc +1|t+1 . x ˆ . let ξ xt+Nc |t+1 − xt+Nc |t . κf (x )) ∈ B(f (x .t+Nc −1|t] is the solution of the FHOCP.6.1. ˆ ˆ it follows that f (x. κf (x )).t+Nc −1|t+1] = [u◦ . Moreover. Lemma 5. suitably chosen to Now. for all x ∈ XNc ¯ ⊕ B(dκf ). with κf (x ) ∈ U .1. u ¯ t+1|t t+N . it follows that by Lemma 5. Robust Open-loop NMPC x ∈ X Nc |x − x | ≤ ε/Lfx . . then. proving that. ε) ⊆ XN . . thus proving the statement.1 and 5. there based on t + 1. u◦ [t.t+Nc −1|t] . ˆ u◦ . . u = u(ˆt+Nc |t+1 ) ∈ U ¯ ¯x is a feasible control action. . ˆ ˆ 1−N ¯ ¯ µ(Υsup )≤Lfx c dκf .100 Chapter 5.2.5.4-5. . with dt ∈ Dt .

6. . .t+Nc −1|t] and u[t+1. ˆ ˆ fx for all for all k ∈ [t + 2. ξ ∈ B(ˆk|t ). ˆ it xk|t +ξ = xk|t+1 +η ∈ X . ρ ˆ xk|t ∈ Xk|t ˆ X η ∈ B(ˆk|t+1 ). t + Nc − 1]. ˆ and hence. 5. . ˆ xt+1|t+1 = f (xt . . . it follows that. ˆ ˆ under Assumption 5. ∀η ∈ B(ˆk|t+1 ). t + Nc − 1]. made respectively using the input and initialized with that sequences and with ¯ u◦ [t. ρ Since ˆ xk|t ∈ Xk|t . it follows that ρt+1|t − ρt+1|t+1 = Lδ |xt | + µ.1. Assuming ˆ ρk|t ˆ given by (5.t+Nc −1|t+1] . consider the predictions xk|t ˆ and xk|t+1 .2 and (5.9). .13) k ∈ [t + 2.5.9). . . in view of Assumptions 5. . k ∈ [t + 2. .13). let introduce Then. ¯ in view of . Appendix Step 1: ˆ xt+j|t+1 ∈ Xt+j|t+1 : ˆ 101 First. . with Now. ˆ ˆ ρ which nally yields ˆ xk|t+1 ∈ Xk|t+1 . . |ξ| ≤ |ˆk|t+1 − xk|t | + ρk|t+1 ≤ Lj−1 (Lδ |xt | + µ) + ρk|t+1 .1. let it follows that ξ xk|t+1 − xk|t + η . . κM P C (xt )). . xt|t = xt ˆ B(ˆk|t ). ρk|t − ρk|t+1 ˆ ˆ ≥ (Lfx + Lδ ) which yields k−t−1 − Lδ (Lfx + Lδ ) k−t−2 k−t−2 j=0 Lfx Lfx +Lδ j (Lδ |xt | + µ) ¯ k−t−1 ρk|t − ρk|t+1 ≥Lfx ˆ ˆ (Lδ |xt | + µ). ρ Furthermore. ˆ k ∈ [t + 2. ˆ Step 2: xt+Nc +1|t+1 ∈ XNc : ˆ if N ¯ Lfxc −1 (δ(|xt |) + µ) ≤ dκf . . . x ˆ ˆ ¯ ˆ fx it turns out that follows that In view of (5. t + Nc − 1] Proceeding by induction. t + Nc − 1]. |ξ| ≤ ρk|t . ˆ ˆ ¯ and ρk|t − ρk|t+1 = (Lfx + Lδ ) ρk−1|t − ρk−1|t+1 + Lδ (|ˆk|t | − |ˆk|t+1 |) ˆ ˆ x x ˆ ˆ ¯ ≥ (Lfx + Lδ ) ρk−1|t − ρk−1|t+1 − Lδ Lk−t−2 (Lδ |xt | + µ). ¯ for all (5.

1) under the action of the Theorem 5. Np ) = t+Np −1 k=t l(ˆk|t . ˆ u ∈U .7).t+Nc −1|t] .15) xt ∈ X M P C . u)∈XNc .3. x x κf (ˆt+Nc −1|t ) x . Nc . it follows that xt+Nc |t+1 ∈ P(XNc ).3: V (xt ) ≥ l(|xt |) for all (5. in view of Assumption 5.6 it follows from M P C is a RPI set for system (5. u◦ [t.2 there exists a feasible control action such that the statement holds.10 is an ISS-Lyapunov function in the set V dened X MP C. . there exists a feasible control action. Then X M P C ⊇ XNc ⊇ Xf . .2: In view of Assumptions 5.102 Chapter 5. sequence for the FHOCP is given by ˜ u[t. ¯ such that Hence.1-5.1 that X control law (5. u[t. .2. 1 .14) for all xt ∈ Xf . Robust Open-loop NMPC Lemma 5. the proof consists in showing that function in 5. in view of Theorem 5. If ¯ ¯ ¯ dκf < LNc −1 (δ(|xt |)+ µ) ≤ d.4 it holds that t+Np −1 ˜ V (xt ) ≤ J (xt . the The very special case Lfc . Now.5.t+Nc −1|t] κf (ˆt|t ). Lfx = 1 can be trivially addressed by a few suitable modications to the proof of Theorem 5. κf (ˆt+1|t ).t+Nc −1|t] . So. ˆx xt+Nc +1|t+1 = f (ˆt+Nc |t+1 .1. ˆ ¯ Proof of Theorem 5. fx thanks to Lemma 5. Suppose 1 that Lfc and Lfx = 1. by Assumption 5. κf (ˆk|t )) + Vf (ˆt+N |t ) x x x p ≤ k=t Vf (ˆk|t ) − Vf (ˆk+1|t ) + Vf (ˆt+N |t ) x x x p ≤ Vf (ˆt|t ) ≤ LVf |xt | x (5. for any xt ∈ XNc . Then. . In fact. The lower bound on V (xt ) can be easily obtained using Assumption 5. given the optimal control sequence at time t. namely thus ending the proof. a feasible control X MP C is not empty. First.

4. . u) − l(ˆt+Nc |t . u◦ ) ≤ Ll Lk−t−1 (δ(|xt |) + µ(|υt |)). . Np ) = V (xt ) − l(xt . . ¯ ¯ f fx fx (5. if L1−Nc dκ < s ≤ L1−Nc d.16) k ∈ [t + 1. x x k|t k|t fx for all (5.3. u◦ ) + t. u◦ ) − l(ˆk|t . t + Np − 1].1 and 5. (u. . κf (ˆt+N x p |t+1 ))) −Vf (ˆt+N x ). t + Nc − 1]. . u) x x x ¯ k|t k|t l(ˆk|t+1 .5. 5. κf (ˆk|t )) x x x x p |t+1 −l(ˆt+Nc |t . u= ¯ u ∈ U : f (ˆt+N ˆx ¯ c |t+1 . u) ∈ XNc .17) where 0. u◦ c −1|t . for all results k ∈ hold [t + Nc + 1. Appendix u◦ . . it follows that l(ˆk|t+1 . under Assumptions the 5. with cost ¯ J (xt+1 .t+Nc |t] with κ (ˆ f xt+Nc |t ). κf (ˆt+Nc |t )) + x x +l(ˆt+N x p |t+1 p |t k=t+(Nc +1) . Using Assumptions 5. following intermediate . Nc .t+Nc |t+1] .18) Finally. . ¯ 1−N ¯ if δ(|xt |) + µ(|υt |) ≤ Lfx c dκf ¯ ¯ if L1−Nc dκ < δ(|xt |) + µ(|υt |) ≤ L1−Nc d fx f fx is a feasible (in general. κf (ˆt+N x p |t+1 ˆx )) + Vf (f (ˆt+N . w) ∈ U × U }.t t+Nc −1 k=t+1 t+Np −1 l(ˆk|t+1 . . κf (ˆk|t+1 )) − l(ˆk|t . u ¯ t+1|t t+N 103 sequence ¯ u[t+1. Moreover. κf (ˆt+Nc |t )) x ¯ x x N ≤ Ll Lfxc −1 (δ(|xt |)+ µ(|υt |)) +Llu ∆u (δ(|xt |) + µ(|υt |)) (5.3 and 5. . l(ˆt+Nc |t+1 . . ¯ if s ≤ L1−Nc dκf fx ∆u (s) max{|u − w|. . u◦ ) − l(ˆk|t .1. u[t+1. .6. for k = t + Nc . suboptimal) control sequence for the FHOCP at time t + 1. u◦ ) + l(ˆt+Nc |t+1 .

for all υ ∈ MΥ .19) Vf (ˆt+Np |t+1 ) − Vf (ˆt+Np |t ) x x ≤ LVf Lfcp where N −(Nc +1) ∆x (δ(|xt |)+µ(|υt |))+LNc (δ(|xt |)+µ(|υt |)) fx (5.21). Np ) it follows that V (xt+1 ) − V (xt ) ≤ −α3 (|xt |) + σ(|υt |). u[t+1. κf (ˆt+N p|t+1 )) + Vf (ˆt+Np +1|t+1 ) − Vf (ˆt+Np |t+1 ) x x x x +LVf Lfcp ≤V N −(Nc +1) Nc Lfx (δ(|xt |) + µ(|υt |)) (xt ) − l(xt . 5. κf (ˆk|t+1 )) − l(ˆk|t . Nc . (5.20) and ∆x (s) = 0 if ¯ s ≤ L1−Nc dκf fx ¯ ∆x (s) = max {|x−ξ|. Now. u◦ ) + ϕx (|xt |) + ϕυ (|υt |) .22) . if hold ¯ δ(|xt |) + µ(|υt |) ≤ L1−Nc dκf .16)-(5.5 and 5.t Np −1 Nc j=1 Ll Lj−1 (δ(|xt |) + µ(|υt |)) fx j−(Nc +1) + j=Nc +1 (Ll + Llu Lκf )Lfc LNc (δ(|xt |) + µ(|υt |)) fx + l(ˆt+N p|t+1 . u◦ ) + t. Nc . N −N −1 and all υ ∈ MΥ . and all Consider now the case ¯ δ(|xt |) + µ(|υt |) ≤ dκf . Assumptions the following inequalities 5.104 Chapter 5. using (5. where ϕυ (|υt |) is a Ll Lfx −1 + (Ll + f K -function. c L p −1 Llu Lκf ) fcLf −1 LNc fx c N −(Nc +1) Nc +LVfLfcp Lfx µ(|υt |) from inequality ¯ V (xt+1 ) ≤ J (xt+1 . fx fx xt ∈ X .t LNc −1 x for all xt ∈ X M P C . Np ) ≤ V (xt ) − l(xt . κf (ˆk|t )) ≤ x x x x (Ll + Llu Lκf )Lfc and k−t−(Nc +1) ∆x (δ(|xt |) + µ(|υt |)) + LNc (δ(|xt |) + µ(|υt |)) fx (5.t+Nc |t] . ξ) ∈ XNc ×(XNc ⊕ B(d))} − LNc s fx if (5.t+Nc |t] .21) ¯ ¯ L1−Nc dκf < s ≤ L1−Nc d. Then. fx ¯ J (xt+1 . u[t+1. (x. t.4. Robust Open-loop NMPC l(ˆk|t+1 .9.

Nc . By (5. only ISpS can be guaranteed.14). fx fx the following inequality can be straightforwardly obtained V (xt+1 ) − V (xt ) ≤ −˜ 3 (|xt |) + σ (|υt |) + σ . hold.t+Nc−1|t .5.23) α3 (|xt |) α3 (|xt |)− (Ll + Llu Lκf ) ˜ Lfcp N −(Nc +1) −1 Lfc − 1 Lfcp N −(Nc +1) + LVf Lfcp N −(Nc +1) LNc δ(|xt |) fx σ (|υt |) ˜ σ(|υt |) − (Ll + Llu Lκf ) L −1 Lfc − 1 −1 + LVf Lfcp N −(Nc +1) LNc µ(|υt |) fx Np −(Nc +1) c (Ll +Llu Lκf ) fc Lf −1 +Lhf Lfcp ) max{|x−ξ|.Np ) M P C . α ˜ ¯ where (5. (u. Nc α3 (|xt |) Np xed. the optimal cost J (xt . Hence. xt ∈ X M P C . although the invariance of X MP C and the fulllment of constraints are preserved thanks to Theorem 5. w) ∈ U ×U}. if such that K∞ -function. then it is a stability margin for the Therefore. δ(|xt |) is where with and for all υ ∈ MΥ .6.7 hold. Conversely.15) and (5. in view for the closed-loop system in X of Theorem 2. Since Assumption 5. ξ) ∈ c ¯ XNc ×(XNc ⊕ B(d))} + Llu max{|u−w|. (x.22).9 is an ISS-Lyapunov function closed-loop system [Jiang & Wang 2001]. if Assumption 5. Appendix α3 (|xt |) l(|xt |)−ϕx (|xt |) and is a 105 σ(|υt |) ϕυ (|υt |). (5. in the latter and σ ¯ N −(N +1) case. if ¯ ¯ L1−Nc dκf < δ(|xt |) + µ(|υt |) ≤ L1−Nc d.u◦ t.1. . it is possible to conclude that the closed-loop system is regional ISS in X MP C with respect to v ∈ Υ.1.

.

. . . . . . . . . . . . . . . . 117 121 126 127 Conclusions . . . . . . . . . . . . . . . . . .1 6. . . .4 Introduction . . . . . . . . . . . .2 6. . . . . .4.1 Introduction The design and analysis of decentralized control systems have been under investigation for more than thirty years. . . . . . . . . . . . .Chapter 6 Decentralized NMPC: an ISS approach Contents 6. . . . .4.3 6. . . . . .6 Open-loop formulation . . Many problems falling into this category have been addressed with various mathematical tools. . . . . . . . . . . . . . . 6. . . 6. . . . . while new theoretical and application issues are arising as a result of current trends . . . .5 6. . . . . . . . .2 6. . . . . 107 109 112 116 Problem statement Regional ISS for interconnected subsystems Nonlinear Model Predictive Control . . . . . . . . . Appendix . . . . . .1 6. . . Closed-loop formulation . . . . . . . .

in fact a decentralized control structure is often the most appropriate one due to topological constraints and limited exchange of information between subsystems. other investigations focused on the structure and size of interconnections [Siljak 1978]. while reducing the computational load stemming from a centralized approach. decentralized MPC techniques are considered. decentralized MPC has already been studied for discrete time linear systems in e. Starting with the notion of xed modes introduced in the 1970s for linear large scale systems [Wang & Davidson 1973]. which is used in several applications. Studies on topology independent control have also been recently reported [Cogill & Lall 2004]. Decentralized MPC is of paramount interest in the process industry. 2002] and in a number of papers quoted there. [Camponogara et al. Works on the size of interconnections have been proposed in adaptive control [Ioannou 1986]. [Dunbar & Murray 2006]. such as ight formation and distributed sensors. where closed loop stability of the origin is achieved through the inclusion of a contractive constraint (see also [de Oliveira Kothare & Morari 2000]). while the MPC approach allows one to include in the problem formulation both performance requirements and state and control constraints. [D'Andrea & Dullerud 2003]). one of the key objectives is to nd conditions guaranteeing closed-loop stability. For these reasons. Specic emphasis on the structural properties of decentralized controlled large-scale systems is given in the research work of D'Andrea and co-workers (see. and the need to come up with more autonomous systems. the availability of spatially distributed sensors and actuators.g. When dealing with large scale systems. In this chapter.108 Chapter 6. Decentralized NMPC: an ISS approach in distributed systems. and more recently in model predictive control. such as the increasing size and complexity of feedback control systems. . for instance. In [Magni & Scattolini 2006] a decentralized MPC algorithm for nonlinear systems has been proposed.

This second possibility is obviously more critical than the previous and requires a more conservative solution. discrete-time systems are derived under the assumption that no information is exchanged between subsystems. and by considering the eect of interconnections as a perturbation term. In this chapter. 6.2 Problem statement Assume that the plant to be controlled is composed by the interconnection of S local subsystems described by the following nonlinear. A similar approach has also been taken in [Dashkovskiy et al. 2005]. 2007] and [Dashkovskiy et al. 2005] where global results are given for interconnected systems. stabilizing fully decentralized MPC algorithms for nonlinear. [Dashkovskiy et al. However this setting more closely resembles most of real world cases. where complex control structures are built according to fully decentralized schemes. as in [Magni & Scattolini 2006]. or by considering a fully decentralized control structure.2. it is shown that under suitable assumptions the ISS property of the controlled subsystems guarantees the ISS of the overall (controlled) system. Rely- ing on the concept of regional ISS. as in [Dunbar & Murray 2006]. 2007]. 2006a]. Problem statement 109 Distributed MPC algorithms can be developed either by assuming that there is a partial exchange of information between the subsystems. 2005]. [Dashkovskiy et al. [Venkat et al. since the amount of information available to any local controller is less. by suitably combining and extending the results reported in [Magni et al. each one of them controlled by a robust (open-loop or closed-loop) MPC algorithm guaranteeing ISS. discrete-time . the approach taken in the following to derive decentralized MPC implementations consists in considering the overall system as composed by a number of interconnected subsystems.6. Then.

sup with Di known.3) Di is a compact set containing the origin as an interior point. it is assumed that the origin is an equilibrium point. . f (x. u1 ). Dening .1 1. which depends on the overall state S xk [x1k x2k .110 Chapter 6. For simplicity of notation. . and gi : IRn → IRni . uik ∈ IRmi is the current control vector. . u) d [f1 (x1 . 0) = 0.e. fS (xS . . uS )] g(x) [g1 (x). . . gS (x)] and [d1 . uk ) + g(xk ) + dk . dS ] . . the whole system can be written as xk+1 = f (xk . k ≥ 0 where (6. uik ) + gi (xk ) + dik . fi (0. di is such that 2. .2) Each subsystem is supposed to fulll the following assumptions. k ≥ 0. . Assumption 6. . . The disturbance di ∈ Di where (6. n i=1 n ni describes the inuence of the S subsystems on the i-th subsystem. (6. . Decentralized NMPC: an ISS approach models xik+1 = fi (xik . 3. . . xSk ] ∈ IR . dik ∈ IR ni is an unknown disturbance.1) xik ∈ IRni is the state of the i-th subsystem. The state and the control variables are restricted to fulll the following . i.

. by point 6 of Assumption 6.}. . . u) − fi (b. There exist positive constants Lij . 7. u)| ≤ Lif |a − b| for all (6. Let denote set on the state. The map fi : IRni × IRmi → IRni is Lipschitz in xi in the domain Xi × U i . .e. .e. The state of the plant In the following. is such that 5.4) (6. there exists a positive constant Lif such that |fi (a. the inter- gi is bounded. i. S] S |gi (x)| ≤ j=1 for all Lij |xj | xi ∈ Xi . . gi ∈ MGi with Gi gi (X sup ). .2. let denote a generic interaction sequence as gi {gi (x1 ). .6) a. both containing the origin as an in- terior point. . 2. j ∈ [1.6. action sequence Note that. i. The map gi : IRn → IRni gi (0) = 0. XS ] the overall constraint 4.5) Xi and Ui are compact sets.1. i. . Problem statement 111 constraints xi ∈ Xi ui ∈ Ui where (6. b ∈ Xi and all u ∈ Ui . such that 6. x ik can be measured at each sample time. X [X1 X2 . gi (x2 ).

dik ∈ IRni is an unknown disturbance. . . FS (xS )] g(x) [g1 (x). without taking into account explicitly the interaction with the other subsystems.7) Fi : IRni → IRni is nonlinear possibly discontinuous. . . a control law ui = κ(xi ). In the following section it is presented a suitable framework for the analysis of stability of such class of closed loop systems: the regional ISS for nonlinear interconnected subsystems. . . subsystem. k ≥ 0 where the map (6. for each local subsystem. Decentralized NMPC: an ISS approach The control objective consists in designing. x ik ∈ IRni and is the state of the i-th subsystem. an optimal closed-loop performance according to certain performance index. . Denand F (x) [F1 (x1 ). discrete-time models xik+1 = Fi (xik ) + gi (xk ) + dik . gS (x)] d . xSk ] ∈ IRn . gi : IRn → IRni .112 Chapter 6. such that the overall system is steered to (a neighborhood of ) the origin fullling the constraints on the input and the state along the system evolution for any possible disturbance and yielding. . . n i=1 ni i-th describes the inuence of the ing S subsystems on the . xk which depends on the overall state S [x1k x2k . 6. . for each subsystem.3 Regional Input-to-State Stability for nonlinear interconnected subsystems Consider a system composed by the interconnection of S local subsystems described by the following nonlinear.

dS ] 113 . .8) The transient of the system (6. . . . Regional ISS for interconnected subsystems [d1 . The state variables xi fulll the constraint xi ∈ Xi where (6. (6. xi .3. 2.9) Xi is a compact set containing the origin as an interior point. disturbance ¯ is denoted di and interaction sequence gi {gi (x1 ). The map gi : IRn → IRni is such that gi (0) = 0. The origin of the system is an equilibrium point. ¯ Each subsystem is supposed to fulll the following as- sumptions. i.7) with initial state sequence by xi0 = xi . the whole system can be written as xk+1 = F (xk ) + g(xk ) + dk .6. Let denote state. . . di .10) k ≥ 0. where Di is a compact set containing the origin as an sup interior point. . k ≥ 0. . XS ] the overall constraint set on the 4. Assumption 6. with Di known.2 1. X [X1 X2 . gi (x2 ).} xi (k. 3.e. Fi (0) = 0. The disturbance di is such that di ∈ Di for all (6. . . gi ).

114 Chapter 6.13) xi ∈ Ξi . with gi (x) + di as disturbance 1.7) is continuous at di = 0 with respect to initial conditions. that means Each subsystem (6. i.2. (6.3 terms. . there exist a suitable (6. satisfying Assumption 6. there exist a suitable a K∞ -function ρi di ∈ MDi ρi such that (id − ρi ) is K∞ -function) and a suitable constant cθi > 0. admits an ISS-Lyapunov function Vi in Ξi ⊆ Xi . Decentralized NMPC: an ISS approach Lij . S] S 5.11) (6. σi such that ∆Vi (xi ) ≤ −αi3 (|xi |) + for all S x j=1 σij (|xj |) d + σi (|di |). . and a pair of suitable K∞ -functions αi1 . given a disturbance sequence and an interaction sequence gi ∈ . all xj ∈ Xj . Ξi is a compact RPI set including the origin as an interior point 2.7). There exist positive constants such that |gi (x)| ≤ j=1 for all Lij |xj | xi ∈ Xi . The solution of (6. ∀xi ∈ Ωi 3. j ∈ [1. . 2. xi = 0. interactions and disturbances. αi2 Vi (xi ) ≥ αi1 (|xi |). ∀xi ∈ Ξi Vi (xi ) ≤ αi2 (|xi |). and all di ∈ Di (with 4. there exist a compact set Ωi ⊆ Ξi (including the origin as an interior such that point).12) K∞ -function αi3 and some suitable K-functions x d σij . such that. gi = 0 ¯ and 6. . Assumption 6.

. . Remark 6. . Lemma 6. there exist some K∞ -functions ai1 .. let dene the maps ∆ : IRS → IRS ≥0 ≥0 ∆(s1 . ai2 . Moreover. (id + aiS )(θS ). with i αi4 S x j=1 σij (||xj ||) d + σi (||di ||)) (6.11) and (6. Note that. gi ) where xi : V (xi ) ≤ bi ◦ ρ−1 .6. (id + ai2 )(θ2 ). in order to verify that and all Θi (di . gi ) ⊆ IΩi {xi : xi ∈ Ωi .16) . .1 Given any K∞ -function ρi such that . . (id + aid )(θd )}.×ΞS In order to introduce the next theorem. Ξ1 ×.15) Let introduce now an intermediate result useful in the following. Regional ISS for interconnected subsystems MGi .1 sequences Note that. 115 there exists a nonempty compact set Θi (di .. . . (id + αS )(sS )) (6. i . . . . . . let dene Ξ the composition of RPI sets of all the subsystems. function Vi is continuous at the origin. gi ) ⊆ IΩ for all di ∈ MDi gi ∈ MGi . . . . by (6. sS ) as ((id + α1 )(s1 ). one has to verify that Θi {x : Vi (x) ≤ bi sup S x j=1 σij (Xj ) sup d + σi (Di )) } ⊆ IΩ (6. (id−ρi ) is a K∞ -function such that too.3. .12). ..14) bi −1 αi4 αi3 ◦ −1 αi2 . |xi |δΩi > cθi } follows (including the origin as an interior point) dened as Θi (di . rst. aid ρ−1 (θ1 + . . + θS + θd ) ≤ max {(id + ai1 )(θ1 ).

18) then the overall system (6.2 and 6.18) together with point 6 of Assumption 6. Theorem 6. sS ) S j=1 −1 x α14 ◦ (id + a1j ) ◦ η1j (sj ).1. Notably. .2 is necessary to guarantee that the interconnections between the subsystems do not cause instability. Let as in (6. 2005].1 Consider systems (6. .7) and suppose that Assumptions 6. such that Γ be given by (6. Decentralized NMPC: an ISS approach αi ∈ K∞ .18) is the generalization to nonlinear interconnected systems of the well known small gain theorem. the results derived in Theorems 2. . while (id + aij ) are obtained starting ρi in (6. x ηij = −1 x σij ◦ αi1 . . .4 Nonlinear Model Predictive Control In this section. If there exists a mapping ∆ (Γ ◦ ∆)(s) s. 2005]. .17) where from αi4 = −1 αi3 ◦ αi2 .1 are used to analyze the ISS property of open-loop and closed-loop min-max formulations of stabilizing MPC for nonlinear systems. in the following it is not . Remark 6.8) is ISS in d to x. which. .116 Chapter 6. Note also that condition (6. .2 As discussed in [Dashkovskiy et al.1 and 6.14) using Lemma 6. ∀s ∈ IRS \ {0} ≥0 Ξ from (6.17).3 are satised. 6. . . Many interesting interpretations of this conditions are given in [Dashkovskiy et al. i = 1.16). . . 1994]. in the case of only two interconnected systems was previously given in [Jiang et al. condition (6. S j=1 −1 x αS4 ◦ (id + aSj ) ◦ ηSj (sj ) with and Γ : IRS → IRS ≥0 ≥0 as (6. S Γ(s1 .

4). the terminal penalty Xif . (6. the constraints (6. let with t1 ≤ t2 ≤ be the with t3 a control sequence.1) with xit = xi . uit3 |t1 ]. the stage cost li . given xik+j|k ˆ predicted state at initial condition k+j obtained with the nominal model fi (xik . ¯ Vif Given and the Ni . Then. Ni ) x k=t subject to li (ˆik|t . the following nite-horizon optimization problem can be stated.6.1 (FHOCP) the positive integer terminal set Consider system (6.5). xik+1 = fi (ˆik . rst let ui[t2 . the performance index t+Ni −1 Ji (¯i . the Finite Horizon Optimal Control Problem (FHOCP) consists in minimizing. with respect to ui[t. uik ) xk and input ui[k. k ≥ 0. . ˆ x k ∈ [t. t + Ni − 1] xt+Ni |t ∈ Xif .t+N −1|t] . Moreover. . uik ). ui[t. 6.t3 |t1 ] [uit2 |t1 uit2 +1|t1 . j ≥ 1. uik|t ) + Vif (ˆit+Ni |t ) x x 1.t+Ni −1|t] . Denition 6.4. the terminal state constraints The stage cost denes the performance index to optimize and satises the following assumption.k+j−1|k] . ˆ with xit = xi ˆ ¯ 3.1 Open-loop formulation In order to introduce the MPC algorithm formulated according to an openloop strategy. the nominal state dynamics 2.4. . Nonlinear Model Predictive Control 117 necessary to assume the regularity of the value function and of the resulting control law.

u)| ≤ Lil |a − b| for all a. 0) = 0 Assumption 6.5).t+N −1|t] .19) Dene the overall control law u κM P C (x1 ) . ui ) ≥ αil (|xi |) xi . .20) Although the FHOCP has been stated for nominal conditions. given x it = x i .e. κM P C (x2 ) . dene κM P C (¯i ) x i where uot|t (¯) x i and apply the control law uot|t (¯) x i is the rst column of uo i[t. i (6. it is possible to guarantee the ISS property of the closed-loop system formed by (6.118 Chapter 6. Lil such that αil is a K∞ -function. i.5 The solution of closed-loop system formed by (6. ¯ nd the optimal control sequence uo i[t. li (xi .19). under suitable assumptions and by choosing accurately the terminal cost function Vif and the terminal constraint Xif . b ∈ Xi and all u ∈ Ui .4 Lipschitz in The stage cost where is such that and is li (xi . u) − li (b. Then. u) li (0. Moreover. ui = κM P C (x). It is now possible to dene a prototype of a nonlinear MPC algorithm: at every time instants t.t+Ni −1|t] by solving the FHOCP. Assumption 6. Decentralized NMPC: an ISS approach li (x.3)-(6. . . there exists a positive constant |li (a. subject to constraints (6. . κM P C (xS ) 1 2 S (6.1) and (6. according to the Receding Hori- zon (RH) strategy.1). . ui ) in the domain Xi × U i .

Xif ⊆ Xi . κif (xi ))) − Vif (xi ) ≤ −li (xi . κif (xi )) ∈ Xif . xi ∈ Xif K∞ -functions αVif and 4. let XiM P C (Ni ) denote the set of states for which a solution of the FHOCP problem exists. gi = 0 ¯ 119 (6. In the following.6. there exist a pair of suitable βVif such that αVif < βVif and αVif (|xi |) ≤ Vif (xi ) ≤ βVif (|xi |) 5. For each ˜ xit ∈ XiM P C (Ni ). Assumption 6. is an admissible. κif (xi )). Assumption 6. given an auxiliary control law κif .4. for all xi ∈ Xif for all fi (xi . closed. 6.t+Ni −1|t] κif (ˆit+Ni |t+1 )] gi ∈ Gi . control sequence for the FHOCP at time di ∈ Di and all possible . possible suboptimal. Xif κif (xi ) ∈ Ui . ui[t+1.19). for all possible [uo x i[t+1. Vif (fi (xi . 2.6 The design parameters Vif and Xif are such that. interactions and disturbances. Vif is Lipschitz in for all xi ∈ Xif Xif with a Lipschitz constant LVif . Nonlinear Model Predictive Control xi = 0. 3. κif (xik )) Xif is asymptotically stable Xif (Vif is a Lyapunov function in for the nominal system).1) and (6.6 implies that the closed-loop system formed by the nominal system fi (xik .t+Ni |t+1] t + 1. Assumption 6.7 Consider closed-loop system (6. 0 ∈ Xif 1.19) is continuous at and di = 0 with respect to initial conditions.

is contained in Remark 6. The sets Let where LiJ LVif LNi −1 + Lil if Lifi −1 Lif −1 . with function Vi given by (6. Ni ) (6. Decentralized NMPC: an ISS approach XiM P C (Ni ) is a RPIA set for the Note that Assumption 6.21) is employed as an ISS-Lyapunov function for the closed-loop system formed by (6. many methods have been pro- posed in the literature to compute Vif .19).6 .8 • Ξi = XiM P C • Ωi = Xif • αi1 = αil • αi2 = βVif • αi3 = αil x • σij = LiJ Lij d • σi = LiJ . dened in (6. Vi (x) Ji (¯i . IΩi .15).t+Ni −1|t] .3 be satised. uo x i[t.1) and (6.1) and (6.7 implies that closed-loop system (6.120 Chapter 6. In what follows.e. Assumption 6.19). Xif satisfying Assumption 6. (depending from N −1 Di and Xj are such that the set Θi sup Di and Xjsup ).21). The but assumptions they are above in can the appear quite dicult of to standard development nonlinear stabilizing MPC algorithms. i. Moreover. the optimal value of the performance index.

2 Closed-loop formulation In the following.1. 6.1. . system (6. X M P C (N ) from d to x. A way to fulll it is shown in [Limon et al. .2 function in Under Assumptions 6.8.1) and (6.19). .9 ists a mapping Given the systems (6.4. However. based on the FHOCP MPC sumption 6.×XS P C (NS ) as the vector of RPI sets of all the subsystems. .5). so as to guarantee that the inuence . Nonlinear Model Predictive Control 121 (see for example [Keerthi & Gilbert 1988. Magni et al. S . (6.1). As- Chen & Allgöwer 1998. 1998c.19) is ISS from gi (x) + di to with RPIA set XiM P C (Ni ). Theorem 6.16). i = 1. In this framework. 2001a]). 6. there ex- Assumption 6.3 (6.20) is ISS in Under Assumptions 6. with the De Nicolao et al. Theorem 6.3)-(6.4-6. (6. dened above. The main result can now be stated. ∆ as in (6.2).9 the overall system (6. it is shown that the ISS result of the previous section is also useful to derive the ISS property of min-max MPC.6. Vi is an ISS-Lyapunov Ξi xi for the closed-loop system formed by (6.1). 2002a] by properly restricting the state constraints 2 and 3 in the formulation of the FHOCP. . 6. Mayne & Michalska 1990. Dene X M P C (N ) M M X1 P C (N1 )×.19) sub- ject to constraints (6.4.18) is satised.4-6. at any time instant the controller for the i-th subsystem chooses the input ui as a function of the current state xi . such that condition (6. that means. .7 is not a-priori satised.

t+Ni −1] and maximizing wi[t.1) with x it = x i . Denition 6. κi[t. the sum of the interaction with the other subsystems and the disturbance is restricted to fulll the following wi where {gi (x) + di } ∈ Wi . Hence. ∀di ∈ Di (6. Ni ) x t+Ni −1 k=t subject to: {li (xik . . with respect to with respect to κi[t.2 (FHCLG) the positive integer and the terminal set Consider system (6. κiNi −1 (xit+Ni −1 )] minimizing the cost in the worst case. containing the origin as an interior point. Decentralized NMPC: an ISS approach S of the disturbance of the subsystems are compensated.t+Ni −1] = [κi0 (xit ) κi1 (xit+1 ) . the state dynamics (6. ∀x ∈ X .10. Note that. ¯ Given Ni .122 Chapter 6. wi[t.23) 1. .t+Ni −1] the cost function Ji (¯i . uik ) − liw (wik )} + Vif (xit+Ni ) (6. Assumption 6. the stage cost li − liw .22) Wi is a compact set of IRni . in view of Assumption 6.t+Ni −1] . the terminal penalty Vif Xif .10 constraint For each subsystem. at any time has to choose a vector of feedback control policies t the controller κi[t. the Finite Horizon Closed-loop Game (FHCLG) problem consists in minimizing. instead of optimizing with respect to a control sequence. the sets of the knowledge of Wi can be derived in view X and Di .t+Ni −1] .1) . The following optimal min-max problem can be stated for the i-th subsystem.

11 is continuous at The solution of each closed-loop subsystem (6. the feedback control law ui = κM P C (xi ) i is obtained by setting of (6. 1 2 S (6. interactions and disturbances. xit+Ni ∈ Xif . the following assumptions are introduced. 6. . 123 2.t+Ni −1] be the solution of the F HCLG. wi[t. . Assumption 6. κM P C (x2 ) .25) In order to derive the main stability and performance properties associated to the solution of FHCLG. . the constraints (6.24) κM P C (xi ) = κo (xi ) where κo (xi ) is the rst element i i0 i0 κo i[t. βiw (|wi |) = βiw (|gi (x) + di |) ≤ βiw S j=1 Lij |xj | + |di | .t+Ni −1] . where The function liw (wi ) is such that and αiw (|wi |) ≤ liw (wi ) ≤ αiw βiw are K∞ -functions.1). Assumption 6.4.12 βiw (|wi |) . Observe that. . gi = 0 ¯ di = 0 with respect to initial conditions. accord- ing to the RH paradigm.5) 3. κM P C (xS ) ] . t + Ni − 1].10 and 6. (6. Dene the overall control law u [κM P C (x1 ) .1.24) and xi . in view of Assumptions 6.3)-(6.12. the terminal state constraint Letting o κo i[t.6.t+Ni −1] . Nonlinear Model Predictive Control k ∈ [t.

13 implies that the closed-loop system formed by the system (6. 2. Xif ⊆ Xi .13 an auxiliary law The design parameters Vif and Xif are such that. The computation of the auxiliary control law.. Assumption 6. there exist a pair of suitable K∞ -functions αVif and αVif < βVif and αVif (|xi |) ≤ Vif (xi ) ≤ βVif (|xi |) for all 5. κif (xi )) + liw (wi ). .. βiw (|wi |) ≤ βiw S j=1 Lij |xj | + |di | (6. . τi2 . Decentralized NMPC: an ISS approach K∞ -functions τi1 . . of the terminal Remark 6. . κif (xi )) + wi ) − Vif (xi ) ≤ −li (xi .13. xi ∈ Xif and all wi ∈ Wi .26) ≤ βiw ◦ (id + τi1 ) ◦ Li1 (|x1 |) + βiw ◦ (id + τi2 ) ◦ Li2 (|x2 |) + . 0 ∈ Xif xi ∈ Xif for all fi (xi . 3. Xif κif (xi ) ∈ Ui . given κif . and all wi ∈ Wi βVif such that 4. .1) and in ui = κif (xik ) is ISS in Xf (Vf is an ISS-Lyapunov function Xf ). in view of Lemma 6.1 there exist some such that . for all 1.4 penalty and of the terminal inequality constraint satisfying Assumption 6. closed.124 Chapter 6. κif (xi )) + wi ∈ Xif .. . xi ∈ Xif for all Vif (fi (xi . + βiw ◦ (id + τiN ) ◦ LiS (|xS |) + βiw ◦ (id + τid )(|di |). τid and. xi ∈ Xif . Assumption 6.

is contained in Theorem 6.6. with function IΩi . In this regard.27) is employed as an ISS-Lyapunov function for the closed-loop system formed by (6. where it is shown how to compute a nonlinear auxiliary control law based on the solution of a suitable under control. i. Ni ) (6. Vi is an ISS- Lyapunov function in XiM P C (Ni ) for the closed-loop system formed by (6.t+Ni −1] .1) and (6.4. κo x i[t. V (x) o Ji (¯i .24). 6. the optimal value of the performance index.10-6.1) . Assumption 6.14 • Ξi = XiM P C • Ωi = Xif • α1 = αl • α2 = βVf • α3 = αl Let x • σij = βiw ◦ (id + τij ) ◦ Lij d • σi = βiw ◦ (id + τid ).14.e. dened in (6. a solution for ane system is discuss in Chapter 3.27). The sets Di and Xj are such that the set Θi (depending from sup Di and Xjsup ).4.4 Under Assumptions 6. Vi given by (6.1. H∞ problem for the linearized system In what follows. Nonlinear Model Predictive Control 125 is not trivial at all. wi[t.15). 6.t+Ni −1] .

Decentralized NMPC: an ISS approach and (6. ∆ as in (6. 6. (6.24) is ISS from gi (x) + di to xi with RPIA set XiM P C (Ni ). In fact. without this term only ISpS can be proven. (6.25) is ISS in X M P C (N ) from d to x. S . . the stability analysis has been performed by considering the interconnections between the subsystems composing the overall system under control like perturbation terms and by using local MPC control laws with robustness properties.5). The main result can now be stated. . Both open-loop and closed-loop MPC formulations have been studied.23) in order to obtain the ISS.18) is satised.24) subject to constraints (6. . (6.15 exists a mapping Given the systems (6. Specically.126 Chapter 6. there Assumption 6. i = 1. Theorem 6.2). . system (6. 6.5 Under Assumptions 6.5 Conclusions Regional Input-to-State Stability have been used in this paper to study the properties of two classes of decentralized MPC algorithms applied for control of nonlinear discrete time systems. that means. .10-6. such that condition (6.1).1.5 Note that the term liw (wi (k)) is included in the performance index (6. Further research is required to establish the eect of partial exchange of information between subsystems on the stability conditions to be fullled. Remark 6.24).16).3)-(6. see Chapter 3. 6.15 the overall system (6.4.1).

there exists a K∞ -function ζi such that ρ−1 = (id + ζi ) i and K-function γ S γ i=1 ri ≤ max{γ(Sr1 ). . . . . . (6. aid ρ−1 (θ1 + . γ(SrS )}. .6. + θS + θd ) i ≤ max{(id + ζi )((S + 1)θ1 ). all xj ∈ Ξj . . (id + ζi )((S + 1)θd )}.13) it follows ∆Vi (xi ) ≤ −αi4 (Vi (xi )) + for all N x j=1 σij (|xj |) d + σi (|di |). . i .. . Proof of Theorem 6.1: From equation (6. it is obtained that ρ−1 (θ1 + . Appendix 127 6. ∆Vi (xi ) ≤ −αi4 (Vi (xi )) + N x j=1 ηij (Vj (xj )) d + σi (|di |). . . Finally. .6 Appendix Proof of Lemma 6. assume that equation (6. (6. . and all di ∈ Di . observe ρi . (id + aiS )(θS ). it is obtained that there exist some such that K∞ -functions ai1 . (id + aid )(θd )}. .1: that given for any As it is proven in [Jiang & Wang 2001]. γ(Sr2 ). (id + ai2 )(θ2 ).30) . (id + ζi )((S + 1)θ3 ). . .. . Using (6.29) implies (id − αi4 ) K-function. ai2 . .. + θS + θd ) ≤ max {(id + ai1 )(θ1 ).11). . −1 αi4 = αi3 ◦ αi2 . . (id + ζi )((S + 1)θ2 ). .6. + θS + θd ) = (id + ζi ) ◦ (θ1 + .29) xi ∈ Ωi . (6. where is a Without loss of generality.28) Using these observations.

1. there exist βi ∈ KL S such that −1 Vi (xik ) ≤ max{βi (Vi (xit ).32) for all xi ∈ Ψi .4) in [Dashkovskiy et al. k). Θi Ri (ei ) {xi : Vi (xi ) ≤ ei }.7). αi4 ◦ (id + aid ) ◦ σi (||di ||)} (6. and all di ∈ Di . for all xi ∈ Ψi .32) satises AG property and is 0-AS in Ψi and hence. and following the same steps of the Proof of Theorem 4 in [Dashkovskiy et al. . all xj ∈ Ξj . By Remark 3. Now. di ∈ Di . let Chapter 6. x + aij ) ◦ ηij (||Vj (xj )||) (6. . KL-function βi and a K-function γi Vi (xik ) ≤ βi (Vi (xit ). k) + −1 +αi4 ◦ (id + S −1 j=1 αi4 ◦ (id d aid ) ◦ σi (||di ||). . 2007]. Note that Θi ⊂ Ψi . that is there exists Hence. In fact. k) + γi (||di ||).6 [Jiang & Wang 2001].33) xi ∈ Ψi . the region Tψi such that Vi (xi (k)) ≤ ei . one has −1 x Vi (xik ) ≤ max{βi (Vi (xit ).2 and there exist a the compactness of such that Ψi .128 xi ∈ Ωi . Decentralized NMPC: an ISS approach xj ∈ Ξj . Then Vi (xik ) ≤ βi (Vi (xit ). ei > bi ¯ and Dene Ψi {xi : Vi (xi ) ≤ ei = ¯ Ψi in maxRi ⊆Ωi ei }. αi4 ◦ (id + ai1 ) ◦ ηi1 (||V1 (x1 )||). This means that the state will arrive in a nite time.32) instead of (2. . and all di ∈ Di .17). for all all and all where Given ei ∈ IR≥0 . k). ai2 . −1 −1 d x αi4 ◦ (id + aiS ) ◦ ηiS (||VS (xS )||). ∀k ≥ Tψi .2 .1. us- ing (6.16) and (6. for all (6. . all xj ∈ Ξj . By Theorem 2. aiS and aid are K∞ -functions. .31) where ai1 . using (6. and all di ∈ Di . . by using point 7 of Assumption 6. the region is reached asymptotically. . Moreover. αi4 ◦ ρ−1 ( i j=1 x d ηij (||Vj (xj )||) + σi (||di ||))}. using Lemma 6. it can be shown that (6. by using point 7 of Assumption 6. −1 x x ηij = σij ◦ αi1 . ¯ Ψi is a RPI set for the subsystem (6. 2007].

¯ Hence the system (6. AG in Ψi + 0-AS in Ψi is equivalent to ISS in Ψi (see [Sontag & Wang 1996] for the continuous time and [Gao & Lin 2000] for the discrete time case).5: by Theorem 6.3: by Theorem 6.4: and by using (6. gether with Assumption 6. ¯ for all and from (6.19) admits an ISS-Lyapunov function satisfying Assumption 6. Proof of Theorem 6.1.34) xi ∈ Ψi . Hence. Ψi implies U AG in Ψi and LS (see Chapter. for any K∞ -function γ . each subsystem (6.1).4 can be derived from the proof of Theorem 2.. from (6.6. the term S j=1 Lij |xj |+|di | Proof of Theorem 6.4.1. To- (6. the proof of Theorem 6. k) −1 γi (||di ||) = αi1 ◦ 2γi (||di ||).3 holds. Appendix 129 and the compactness of Ψi .7) is ISS in Ψi di to Then.3 by substituting at |w|. considering that.1.11) and (6. the state will reach the region Ψi in a nite time. the term S j=1 Lij |xj | + |di |.7) is ISS in Ξi from di to xi and hence the overall system (6. the system (6. by Lemma 2.2. γ(r + s) ≤ γ(2r) + γ(2s). . xi .3.8) is ISS in Ξ from d to x. where −1 ¯ βi (|xit |.28).26).1 by substituting at |w|. each subsystem (6. k) = αi1 ◦ 2βi (αi2 (|xit |). k) + γi (||di ||). Now. using the properties (6. UAG in Ψi implies UAG in Ξi . Considering that starting from Ξi . Proof of Theorem 6. and using Theorem 6. and all di ∈ Di .12). Proof of Theorem 6. using again Lemma 2. ISS in for the denitions). one has ¯ |xik | ≤ βi (|xit |.6.2: the proof of Theorem 6.9.2 can be derived from the proof of Theorem 2. this proves that Theorem 6.1)..

.15.3. To- gether with Assumption 6. Decentralized NMPC: an ISS approach (6.1.5 holds.130 Chapter 6. this proves that Theorem 6.24) admits an ISS-Lyapunov function satisfying Assumption 6. and using Theorem 6.

. . . . . . but a considerable number of publications in this area focus on specic issues related to Unin- . . For instance. . . . .Chapter 7 Cooperative NMPC for Distributed Agents Contents 7. .3 Introduction . . . . . . . . . . .3. . . . . . . . . . . 7. . . .1 Introduction Another research direction in decentralized control considers the problem of controlling a team of dynamically decoupled cooperating systems.3. . . . . Conclusions . . . . . .2 7. . .1 7. . .5 Regional ISS results . . . . . . . . . . . . . . .3 7. . 141 143 147 153 158 Stability properties of the single agents Stability properties of the team of agents Simulation results . 2003]. . .2 7. . . 131 134 140 Problem statement Stability of the team of cooperating agents . . there have been some important theoretical re- sults on the stability of swarms [Liu et al. . . 7. . . . . . . . . . . . . . . . . . . .1 7. . . .4 7. . . . . .3. . . . . . . . . . . . . . . . . . . . .

where a centralized RH problem is decomposed and solved locally. such as collision avoidance and formation constraints. [Li & Cassandras 2003] and [Li & Cassandras 2004]. 2004] the concept of Leader to For- . The cost function used for MPC framework can take into account several issues. 2002]. Coordination of a large group of cooperating nonlinear vehicles is considered in [Dunbar & Murray 2004] and related works. where a centralized problem is decomposed to allow local computations and feasibility issues are thoroughly examined. 2004a] exploiting a hierarchical decomposition of the team in suitable subgraphs with assigned priorities. for instance. In [Li & Cassandras 2002]. 2006]). 2000. which every agent computes for itself and its neighbors in the model predictive control process. Y. Jin et al. Polycarpou et al. [Tanner et al. the authors consider a two-degrees of freedom team of UAVs assigned to visit a certain number of points. an ISS analysis has recently been proposed by several authors. [Kevicky et al. 2001. One of the approaches used in this area is based on the selection of a suitable cost function and its optimization in a model-predictive control (MPC) framework. and may reward the tracking of a certain path. Cooperative NMPC for Distributed Agents habited Autonomous/Air Vehicles (UAVs) applications (see. 2004b]. [Chandler et al. stability is obtained in [Kevicky et al. the authors prove certain stationarity properties of the generated trajectories in the case of two agents searching for multiple targets. 2004a]. In [Tanner et al. The team of UAVs is controlled in a centralized recedinghorizon (RH) framework and by exploiting global potential functions. Towards a broad analysis of the structural properties of cooperative systems. A RH control scheme has also been proposed in [Kevicky et al.132 Chapter 7. Convergence to the formation equilibrium point is assured by guaranteeing frequent updates and a bounded error between the assumed and the predicted trajectories.

Hence. the derived result conrms that. A rigorous stability analysis is carried out. where Nesic and Teel propose a new unied framework for modelling and analyzing networked control systems [Nesic & Teel 2004b]. Each agent is assumed to evolve in discrete-time by means of locally computed control laws. ISS tools have been successfully applied to the specic case of networked systems with serial communication. The in- formation ow among the agents is considered as a set of interconnections whose size is measured by the weight this information has in the computation of the control action.1. a cooperative control problem for a team of distributed agents with nonlinear discrete-time dynamics is considered. which takes into consideration delayed state information from a subset of neighboring cooperating agents. The asymptotic stability of the team of agents is then proved by utilizing smallgain theorem results. The cooperative control problem is rst formulated in a MPC framework.7. Introduction 133 mation Stability is developed. a suitable interconnection boundedness is necessary to guarantee stability. In this chapter. 2004] to a nonlinear framework. which is analyzed using an ISS approach. exploiting the regional ISS properties of the MPC local control laws. in this framework. The problem formulation is based on a completely decentralized MPC control algorithm. The proposed scheme generalizes the approach presented in [Franco et al. A discussion of some of the issues aris- ing in the study of non-holonomic vehicles using ISS can be found in [Chen & Serrani 2004]. . [Nesic & Teel 2004c]. where the control laws depend on the local state variables (feedback action) and on delayed information gathered from cooperating neighboring agents (feedforward action).

M . 0) = 0..134 Chapter 7. i = 1. for each (7.3) In openloop mode. Each agent nonlinear time-invariant state equation: xik+1 = fi (xik . and where it is assumed that dynamics of all fi (0. Let also suppose that the M agents evolve on the same discrete-time space (that is. in Section 7... The state vector xi of each agent Ai : i = 1.2 Problem statement In this section. M is constrained to belong to a compact set Xi . . Cooperative NMPC for Distributed Agents 7.. that is. . the stability properties of the distributed controlled system will be analyzed and the main results will be proved.. . xik ∈ IRni denotes the local state vector and uik ∈ IRmi denotes the local control vector of agent Ai at time k . M } is considered. (7. A distributed dynamic system made of a set of M Ai agents denoted as is described by the A {Ai : i = 1. M . the control vector set (7. ui ∈ Ui ⊂ IRmi . the cooperative control problem addressed in the chapter will be formulated in general terms. k ≥ 0.. Analogously. whereas. The coupling between agents arises due to the fact that they operate in the .. each agent is dynamically decoupled from the remaining agents and the dynamics of the other agents are not assumed to be known.3. xi ∈ Xi ⊂ IRni . that is.... the agents are synchronized).. .2) ui is constrained to take values in a compact Ui .1) i = 1. uik ). xi0 = xi ¯ where..

For each i = 1. To gain that is. . Let us consider a generic timeinstant t.7. routing nodes in communications and/or transportation networks where agents cooperate to minimize the total trac delay. let group all inputs to agent Ai into a vector is equal to wit ¯ dened as wit ¯ col (xit−∆ij . where a simple threeagent example is shown pictorially. Wi Π Xj .. let now introduce . To achieve some degree of cooperation. networks of reservoirs in water-distribution networks.4) Wi j∈Gi denotes the cartesian product of all sets Xj . The size of vector wi ¯ nwi = j∈Gi nj and clearly wi ∈ Wi ¯ where (7. refer to Fig.. etc. agent Ai receives the vector xjt−∆ij from agent Aj ∈ Gi . M and for given values of the state vector xit ∈ Xi and of the information vector wit ∈ Wi ¯ at timeinstant t.2. It is worth noting that the above setting allows the investigation of quite a large class of distributed cooperating dynamic systems like teams of mobile vehicles. Problem statement 135 same environment and due to the cooperative objective imposed on each agent by a cost function dened later on. . 7. j ∈ Gi . the information exchange pattern is dened as follows. j ∈ Gi ) . then for each i = 1. At each timeinstant t. In this specic example. with a given set of neighboring agents Gi {Aj : j ∈ Gi }... cooperating robotic arms. where denotes the set of indexes identifying the agents belonging to the set More precisely. each agent information vector Ai exchanges an wi Gi Gi . .1. some more insight into the information exchange pattern. Aj ∈ Gi the agent Ai receives from each neighboring cooperating agent the value of its local state vector with a delay of ∆ij time steps. . each agent receives information from all remaining agents. M . . that is.

M denote the lengths of the so-called control and prediction horizons.t+Nip ] . wik . dqik k=t The positive integers + Vif xit+Nip .1: Three agents exchanging delayed state information. dli [t. dqi [t. Cooperative NMPC for Distributed Agents Figure 7. 2001a]. . wit . and a cooperation cost term given by t+Nip −1 qi xik .136 Chapter 7. dlit+N ip Nic and Nip . uik . . Nip ) = t+Nip −1 li xik .t+Nip −1] . The local cost function is composed of two terms: a partial cost term given by t+Nip −1 li xik . dqik k=t . respectively. the following nitehorizon (FH) cost function (in general.t+Nic −1] . ui[t. nonquadratic): Ji (xit . . . uik . i = 1. dlik + qi xik . dlit+N k=t where ip li is a transition cost function and Vif is a terminal cost function. according to the framework proposed in [Magni et al. Nic . wik . dlik + Vif xit+Nip .

M . whereas the control variables uik . . t + Nic − 1 will be the argument of a suitable optimization problem. i = 1. . . t + Nip − 1 uik = κif (xik ) . at timeinstant t. . t + Nip − 2 .. let suppose that.g. k = t + Nic . the vectors dl i are useful to specify some reference value for some or all components of the local state variables. for the sake of simplicity. .7.2. . . In the following. . vectors wi t can be considered as known external inputs in the cost function. the control variables uik . whereas the vectors dqi can be used to parametrize the cooperation be- tween the agents. . . . a forward-forgetting-factor is introduced in the cost function as regards the information vector exchanged at time-instant t). . where wit wi ¯ (7. dqi .5) Awi αwi Inwi with αwi < 1 and with Inwi denoting the identity matrix of dimension nwi . For example (see also Section 7. by a suitable change of state coordinates. it is possible to consider an equivalent formulation where the cost function (with straightforward re-denitions of . . then vectors dli . k = t. 137 The quantities denote some given vectors of appropriate dimensions. . The dynamic system (7. Problem statement dli . . if the agents represent UAV vehicles.4). .5) is introduced in or- der to decrease the importance of the information vector in the FH cost function along the prediction horizon (e. dqi could be dened so as to specify given trajectories to be followed by each agent and also given formation structures for the agents. k = t. It is worth noting that. In general. As will be subsequently claried. will be obtained through some auxiliary control law The vector wi denotes the state of the dynamic system wik+1 = Awi wik .

as initial conditions.5) with x it xi ∈ Xi ¯ and wi ∈ W i ¯ Xif . Given the positive integers Nic . wit with i = systems (7. Cooperative NMPC for Distributed Agents the symbols) can be re-written in the simpler form Ji (xit . 2000]). 0) = 0 .t+Nic −1] .4 for some details about the above change of coordinates in a practical simple case). wit . However. . . considering the general case would not involve major conceptual diculties. the Finite Horizon Optimal Control Problem . Vif . The local control law is designed according to a RH strategy. In the literature several dierent problem formulations can be found depending on the particular setting.6) li (0. for every agent Ai . qi (0. .1 Beyond allowing for a simpler problem formulation. the MPC control problem has been stated according to [Magni et al. the origin is an interior point of the sets and Ui .138 Chapter 7. qi . and cooperation and terminal cost functions li . Problem 7. wik ) ] + Vif xit+Nip k=t where (7. Remark 7.6) will allow for the design of time-invariant control laws. uik ) + qi (xik . after a change of coordinates it will also be possible to carry on the stability analysis with reference to the origin as equilibrium state of the time-invariant system (see Section 7. Moreover. Nic .1 (FHOCP) 1. the terminal set the auxiliary control law κif . the transition. Nip ) = t+Nip −1 [li (xik . Nip . In this chapter. the re- duction of the original FH cost function to the form (7. Xi and Vif (0) = 0 .2. ui[t. Consider.1) and (7. 0) = 0 . M . . 2001a] (see also the wellknown survey paper [Mayne et al.

7.2. Problem statement

139

(FHOCP) consists in minimizing with respect to mance index (7.6) subject to:

ui[t,t+Nic −1]

the perfor-

1. the agent's dynamics (7.1) 2. the system's dynamics (7.5) 3. the auxiliary control law

**uik = κif (xik ), k = t + Nic , . . . , t + Nip − 1
**

4. the constraints (7.2) and (7.3) 5. the terminal state constraint

**xit+Nip ∈ Xif . uo i[t,t+Nic −1]
**

solv-

Clearly, by denition, the optimal FH control sequence

ing Problem 7.2.1 is such that, when applied to (7.1), the constraints (7.2), (7.3), and the terminal constraint

xit+Nip ∈ Xif

are simultaneously satis-

ed. Indeed, the following denition regarding a generic control sequence

ui[t,t+Nic −1]

will be useful in the analysis reported in Section 7.3.

**Denition 7.1 (Admissible control sequence)
**

xit ,

the sequence

Given an initial state

ui[t,t+Nic −1]

is said to be an admissible control sequence

for the FHOCP if its application to (7.1) under the action of the auxiliary control law

uik = κif (xik ), k = t + Nic , . . . , t + Nip − 1

allows simultaneous

satisfaction of (7.2), (7.3) and of the terminal constraint

xit+Nip ∈ Xif .

Now, the RH procedure can be described in the usual way as follows. When the controlled agent

Ai

is in the state

x it

at stage

t,

the

FHOCP is solved, thus obtaining the sequence of optimal control vectors,

uo i[t,t+Nic −1] .

ne

Then, according to the Receding Horizon (RH) strategy, de-

κM P C (¯i , wi ) x ¯ i

uot (¯i , wi ) i x ¯

where

uot (¯i , wi ) i x ¯

is the rst element of

140

Chapter 7. Cooperative NMPC for Distributed Agents

uo i[t,t+Nic −1] .

This procedure is repeated stage after stage and a feedback

feedforward control law

**uit = κM P C (xit , wit ) i
**

is obtained, that depends on the local current state of delayed states agents

(7.7)

x it

and on the vector by the cooperating

wit

communicated to the agent

Ai

Gi = {Aj , j ∈ Gi }. RH

optimal control law can

The system (7.1) under the action of the thus be rewritten as

˜ xik+1 = fi (xik , wik )

fi (xik , κM P C (xik , wik )) , k ≥ t, xit = xi ¯ i

(7.8)

It is worth noting that, from well-known results on RH control (see, for instance, [Mayne et al. 2000] and the references cited therein), there is

κM P C (0, 0) = 0

state for agent

and hence when

˜ fi (0, 0) = 0 , that is, the origin is an equilibrium

Ai

wik = 0, k ≥ t . xi = 0 ¯

Assumption 7.1

and

The solution of closed-loop (7.8) is continuous at

wi = 0

with respect to disturbances and initial conditions.

**7.3 Stability of the team of cooperating agents
**

The stability analysis of the team of cooperating agents will be carried out in three main steps. In Subsection 7.3.1 some basic results concerning the regional inputtostate stability properties of discretetime systems will be stated and proved, extending the approach presented in Chapter 2 to this particular case. In Subsection 7.3.2, the regional stability results will be exploited referring to specic dynamic models (7.8) of

Ai .

It will be indeed

7.3. Stability of the team of cooperating agents

141

proven that each agent is regionally ISS with respect to the input represented by the delayed incoming information from its neighbors. Finally the team of cooperating agents will be considered in Subsection 7.3.3 as a single dynamic system resulting from a feedback interconnection of regionally ISS systems. Showing that both the elements of this interconnection are

endowed with ISS-Lyapunov functions, will result in proving the asymptotic stability of the team of cooperating agents by resorting to appropriate smallgain conditions.

**7.3.1 Regional ISS results
**

The regional ISS stability analysis will now be associated to the existence of a suitable Lyapunov function (in general, a-priori non smooth). Note that, dierently from what stated in Chapter 2, in this case, in order to t the considered problem, the ISS-Lyapunov function will be a function of both variables

x

and

w.

Consider a discrete-time autonomous nonlinear dynamic system described by

˜ xk+1 = f (xk , wk ), k ≥ 0,

where the map

(7.9)

˜ f : IRn × IRq → IRn wk ∈ IRq

is nonlinear possibly discontinuous,

xk ∈ IRn

is the state,

is an unknown disturbance. The transient

of the system (7.9) with initial state is denoted by assumptions.

x0 = x ¯

and disturbance sequence

w

x(k, x, w). ¯

This system is supposed to fulll the following

Assumption 7.2

1. The origin of the system is an equilibrium point, i.e.

˜ f (0, 0) = 0.

142

**Chapter 7. Cooperative NMPC for Distributed Agents
**

w

2. The disturbance

is such that

w∈W

where

(7.10)

W

is a compact set containing the origin, with

W sup

known.

Assumption 7.3

The solution of (7.9) is continuous at

x=0 ¯

and

w=0

with respect to disturbances and initial conditions.

**Denition 7.2 (ISS-Lyapunov function in Ξ)
**

IRq → IR≥0

1. is called an ISS-Lyapunov function in

A function

V : IRn ×

Ξ

for system (7.9), if

Ξ

is a compact RPI set including the origin as an interior point

2. there exist a compact set

Ω⊆Ξ

(including the origin as an interior such that

point), and a pair of suitable

K∞ -functions α1 , α2

V (x, w) ≥ α1 (|x|), ∀x ∈ Ξ, ∀w ∈ W V (x, w) ≤ α2 (|x|) + σ1 (|w|), ∀x ∈ Ω, ∀w ∈ W

3. there exist a suitable

(7.11)

(7.12)

K∞ -function α3

and some suitable

K-function

σ2

and

σ3

such that

˜ V (f (x, w1 ), w2 ) − V (x, w1 ) ≤ −α3 (|x|) + σ2 (|w1 |) + σ3 (|w2 |)

(7.13)

for all

x ∈ Ξ,

and all

w1 , w2 ∈ W K∞ -function ζ

and

4. there exist some suitable is a

**ρ (with ρ such that (id−ρ) cθ > 0,
**

such that, given a

K∞ -function)

and a suitable constant

disturbance sequence

w ∈ MW ,

there exists a nonempty compact set

**7.3. Stability of the team of cooperating agents
**

Θw ⊆ IΩ {x : x ∈ Ω, |x|δΩ > cθ }

143

(including the origin as an

interior point) dened as follows

Θw

where

{x : V (x) ≤ b(||w||)}

with

(7.14)

b

−1 α4 ◦ ρ−1 ◦ σ4 ,

α4

α3 ◦ α−1 , α3 (s) 2 ζ(s) + σ2 (s) +

min(α3 (s/2), ζ(s/2)), α2 (s) σ3 (s).

α2 (s) + σ1 (s), σ4 (s)

Remark 7.2

Note that, in order to verify that

Θw ⊆ IΩ

for all

w ∈ MW ,

one has to verify that

Θ

{x : V (x) ≤ b(W sup )} ⊆ IΩ

(7.15)

A sucient condition for regional ISS of system (7.9) can now be stated.

Theorem 7.1

Suppose that Assumption 7.2 and 7.3 hold. If system (7.9)

admits an ISS-Lyapunov function in disturbance sequences

Ξ,

then it is ISS in

Ξ

and, for all

w ∈ MW , limk→∞ |x(k, x, w)|Θw = 0. ¯

To sum up, in this subsection an important sucient condition for regional ISS of constrained systems of the form (7.9) has been stated and proved. In the next subsection, Theorem 7.1 will be exploited with reference to each agent

Ai

under the action of the local MPC control law.

**7.3.2 Stability properties of the single agents
**

Consider a generic agent

Ai

whose dynamics is described by (7.1).

By

exploiting the results proved in Subsection 7.3.1, it will now shown that

li is Lipschitz with respect to and ui in Xi × Ui . . i = 1. Cooperative NMPC for Distributed Agents Ai . with Lipschitz constants denoted as Lliu .. there exist a pair of suitable βVif such that αVf < βVif and αVif (|xi |) ≤ Vif (xi ) ≤ βVif (|xi |) 6. |κif (xi )| ≤ Lκif |xi |. Assumption 7. each agent is considered as a separate dynamic system in the team. where ψi is a K-function and wi ˜ (Awi )Nip −1 wi . 2. . Let now introduce some further useful assumptions and denitions. . over. 2. and all wi ∈ W i .144 Chapter 7.5 li (xi .4 1. Assumption 7. The design parameters Vif and Xif are such that. |fi (xi . M t.e. . Lκif > 0. κif (xi )) − qi (xi . for all The partial cost function and all li ri is such that is a ri (|xi |) ≤ More- xi ∈ Xi . the coupling between the agents is not directly taken into account). 0 ∈ Xif for all Xif ⊆ Xi . Lfic > 0. . 4. for all for all x ∈ Xif xi ∈ Xif xi ∈ Xif K∞ -functions αVif and 5. ui ). κif (xi )))− Vif (xi )≤ i (xi . κif (x) ∈ Ui . Lli ui ∈ Ui xi where and K∞ -function. in the sense that the input vectors wit are external variables that are assumed not to depend on the behavior of the other cooperating agents (i. 3. given an auxiliary control law κif . f (xi . wi ) + ψi (|wi |) −l ˜ ˜ all for xi ∈ Xif . Vif (fi (xi . each agent with is regionally ISS with respect to the inputs represented by the information vectors cooperating agents at each timestep wit received from its Clearly. at the present stage. κif (xi ))| ≤ Lfic |xi |. κif (xi )) ∈ Xif . in this context. respectively. Xif closed.

In what follows.t+Nic −1] . κif (xit+1 ).8 • Ξ = XiM P C • Ω = Xi κif • α1 (s) = ri (s) Suppose 1 that Lfic = 1 and let The very special case Lfic = 1 can be trivially addressed by a few suitable modications to the proof of Theorem 7. In the following. wi ). Assumption 7. Moreover qi is Lipschitz with xi and Xi × Wi . Nip ) (7. .e. Stability of the team of cooperating agents Assumption 7.3.t+Nic −1] col [κif (xit ). . . Nic .1) and (7. with Lipschitz constants denoted as Lqi and Lqiw . let XiM P C denote the set of states for which a solution of the FHOCP problem exists. wit .6 qi (xi .4 are satised. the optimal value of the performance index. 1 .2.1).7). i. uo i[t. Assumption 7. wi in wi ∈ W i .16) is employed as an ISS-Lyapunov function for the closed-loop system formed by (7.7. Vi (xit .7 x it for which Consider system (7. respect to for all 145 0 ≤ The cooperation cost function and all qi is such that xi ∈ X i . κif (xit+Nic −1 )] is an admissible control sequence for the FHOCP and for which points 2 and 3 of Assumption 7. respectively. . Let denote Xi κif the set of states ˜ ui[t. wit ) Ji (xit .

the following useful result is given. .4 holds. Lemma 7. sub- Ai .1.4).2). under slightly more restrictive hypotheses on the agents' dynamic models and on the FH cost function. ui ) = xi Qi xi + ui Ri ui ˜ qi (xi . In this respect. is ISS with RPIA set XiM P C .2).15). a key aspect is the design of an auxiliary control law κif (xi ) such that Assumption 7. . . It is worth noting that.4-7.8). The main result can now be stated. from the perspective of determining regionally ISS stabilizing control laws. (7. M . dened in (7. the locallycontrolled agent whose closedloop dynamics are described by (7. Cooperative NMPC for Distributed Agents (Lfic )Nip −1 s Lfic −1 • α2 (s) = (Lli + Lliu Lκif + Lqi ) • α3 (s) = ri (s) • σ1 (s) = Lqiw (αwi )Nip − 1 s αwi − 1 ip + βVif (Lfic s) N • σ2 (s) = αwi Lqiw • σ3 (s) = Lqiw The set (αwi )Nip − 1 s + ψi ((αwi )Nip −1 s) αwi − 1 (αwi )Nip − 1 s αwi − 1 Θ (depending from W is such that the set W sup ). with function V given by (7.3). 7. and li (xi .1 Assume that fi ∈ C 2 . ject to constraints (7. and (7. Theorem 7.2 Under Assumptions 7. i = 1.8. is contained in IΩ. . wi ) ≤ xi Si xi + ψi (|wi |) .146 Chapter 7.

one can local law solving Problem 7. suppose that there exists a matrix is stable with Ki such that Aicl = Ai + Bi Ki Ai ˜ Qi ∂fi ∂xi . Stability of the team of cooperating agents Qi . M } i = 1.. .7. there exist a constant for all (7.3 Stability properties of the team of agents In this subsection. . ui =0 with ∂fi ∂ui xi =0. let consider the team ating agent A = {Ai . Bi xi =0. Vif = xi Πi xi . ˜ Si ψi 147 K- with and being positive denite matrices and being a function. In the next subsection. Ri . In this respect.. .. and denote by Πi the unique symmetric positive denite solution of the following Lyapunov equation: ˜ Aicl Πi Aicl − Πi + Qi = 0.4 with κif (xi ) = Ki xi .2.1 for each . M . Furthermore. ui =0 Let ˜ βi (Qi + Ki Ri Ki + Si ) βi > 1. Then.3. i = 1. . the coupling eects due to the exchange of the delayed state information between the cooperating agents will be taken into account in the context of the stability analysis of the whole team of agents. 7. the stability analysis of the whole team of agents will be addressed. where each cooper- Ai is controlled by the regionally ISSstabilizing MPC control Therefore.17) Υi ∈ IR>0 Xif and a nite integer ¯ Np such that satises ¯ Np ≥ Np the nal set {xi ∈ IRni : xi Πi xi ≤ Υi } Assumption 7. .3.

First. Cooperative NMPC for Distributed Agents write ˜ x1t+1 = f1 (x1t . wMt ] . wMt ) . w1t ) ˜ x2t+1 = f2 (x2t . M First of all.18) ˜ f (xt . . . i. fM (xMt . w2t ). i = j} . To this end. κM P C (x1t . 1 f2 (x2t . ρτt . are introduced to store the and nz dim (zt ) and where the variables delayed states. ∆ . · · · . wt ) where (7. . . . ∆ zt that is. w2t ) . w1t )) . f2 (x2t . . xMt ] .. wt col [ w1t . M . . let rewrite the team of dynamical systems as a suitable interconnection of two composite systems. ρ Then. . · · · .. · · · . τ = 2. wt ) Vector ˜ ˜ ˜ col f1 (x1t .. M . let introduce the state vector where col [ ρ1t . . For the sake of simplicity and without loss of generality. wt can be easily characterized as the output of a system describ- ing the delay dynamics of the information exchange process among the agents. zt ∈ IRnz . f1 (x1t . specically ρ1t+1 = xt ρτt+1 = ρτ −1t . let xt col [ x1t . w2t )) . 2 ˜ xMt+1 = fM (xMt . κM P C (xMt . . · · · . let set max{∆ij . κM P C (x2t . it is assumed that each agent receives at least one delayed state information from another neighboring agent. wMt ) fM (xMt . j = 1. .148 Chapter 7. i = 1. w1t ). ρ∆t ] . wMt )). · · · . Hence the following state equation can be written: ˜ xt+1 = f (xt . it is assumed that dim (wit ) ≥ 1. .

... ··· ∅ . C = . where nx δij (τ ) I is equal to the identity matrix of dimension nj . ∅ Iτ . otherwise j -th It is δij (τ ) = 0.. thus in matrix the matrix value of τ. i = j only for τ corresponding to the delay associated with the information received by the i-th from the agent. . . . . thus there is no replication of information.7. B = . . ··· ··· . . . . Stability of the team of cooperating agents 149 Hence. . ∅ δi2 (τ ) ··· . . . . . Summing up. for ··· ··· ∅ δiM (τ ) τ = 0. . .3. . .19) = C zt I I1 ∅ · · · · · · ∅ 0 ∅ A = ∅ I2 ∅ ··· ∅ .. ... . j = 1.. . . M . δi3 (τ ) · · · . i. ∅ ∅ · · · · · · I∆−1 ∅ Ci = Ci (τ ) = All matrices ∅ ··· ··· ··· ∅ C1 C2 . Ai C δij (τ ) worth noting that agent does not get replicated information from agent is equal to the identity for only one Aj . it follows that zt+1 = A zt + B xt wt where (7. . .. . . the overall state equation describing the dynamics of the . . ∆ − 1 dim(xt ) and are identity matrices of dimension nx × nx . . ∅ ∅ ∅ . CM Ci (1) · · · Ci (τ ) · · · Ci (∆) δi1 (τ ) ∅ ∅ .

8.2. After this step. In the following. Θ can now be proved. wt ) ≤ −α4 (V (xt . given a suitable K∞ - function as in point 4 of Denition 7.4-7.150 Chapter 7. α2 (s) + σ1 (s). it will be shown that an ISS-Lyapunov function can be dened for each of these systems. wt )) + ζ(|wt |) +σ2 (|wt |) + σ3 (|wt+1 |) for all xt ∈ X κf . respectively. wt+1 ∈ W .44) it follows immediately that the ISS-Lyapunov function V (xt .18) and (7. which implies that both will turn out to be regionally ISS. let W W1 × · · · × WM . where.2 X MP C and Under Assumptions 7. X κf Θ1 × · · · ΘM . The following intermediate result Lemma 7. (see for details the Step 1 of Now. Cooperative NMPC for Distributed Agents team of agents can be written as a feedback interconnection between the systems described by the state equations (7. from (7.18) and (7. wt ) in VD (zt ) in IRnz .1.1). X MP C M M X1 P C × · · · × X M P C . the stability properties of the team of agents will be analyzed by resorting to nonlinear smallgain theorem arguments. α2 (s) the Proof of Theorem 7. from the proof of Lemma 7. X X1 × · · · × XM .19) are provided with suitable ISS-Lyapunov functions V (xt . α3 (s) 2 one has min(α3 (s/2). Now. .19).19) wt = Czt .2. α4 (s) α3 ◦ α−1 (s). since from (7. First. X1 1f × · · · × XMM f κ κ .7. wt ) satises V (xt+1 . wt+1 ) − V (xt . dynamic systems (7. ζ(s/2)). ζ and all wt .

and all w ∈ MW .20) = −α4 (V (xt . wt+1 ) − V (xt . as far as the ISS-Lyapunov function from (7.21) α4D (s) −1 α3D ◦ α2D (s).20) and (7. where σ4 (s) −1 −1 ζ(α1D (s))+σ2 (α1D (s))+ −1 σ3 (α1D (s)). ∀k ≥ Tψ . Since the region Ψ ⊇ Θw and that Θw is reached asymptotically. VD (zt+1 ) − VD (zt ) ≤ −α4D (VD (zt )) + σ4D (V (xt . wt )) + ζ(α1D (||VD (z)[t+1] ||)) −1 +σ2 (α1D (||VD (z)[t+1] ||)) −1 +σ3 (α1D (||VD (z)[t+1] ||)) (7. that is. wt )) + ζ(α1D (VD (zt ))) −1 −1 +σ2 (α1D (VD (zt ))) + σ3 (α1D (VD (zt+1 ))) −1 ≤ −α4 (V (xt . from (7. by using (7. ¯ Ψ is a RPI set.22) . (7. ¯ ˆ βD in [Jiang & Wang 2001]. It is clear that Let Ψ {x : V (x) ≤ e = maxR(e)⊆X κf e.45). Moreover. the state will arrive in there exists Ψ in a nite time. let R(e) {x : V (x. again. k − t).1. wt )) where. wt ). γ1 (||VD (z)[k] ||)} (7.46) it follows that VD (zt ) is concerned. ∀w ∈ W}.7. Analogously to the proof of Theorem 7. wk ) ≤ e.21) it follows that there exist some ˆ KL-functions β and such that ˆ V (xk . Stability of the team of cooperating agents |wt | ≤ |zt | 151 and. wt )) + σ4 (||VD (z)[t+1] ||) for all xt ∈ X κf . wt ) ≤ −α4 (V (xt . −1 V (xt+1 . wk ) ≤ max{β(V (xt .3. given e ∈ IR≥0 . w) ≤ e. Thanks to Remark 3. whereas σ4D (s) −1 σD ◦ α1D (s) . ∀w ∈ W}.7 Tψ such that V (xk .

23) x∈X and all k ∈ Z≥0 . Theorem 7. Then the team of cooperating agents described by the interconnected dynamic equations (7. Cooperative NMPC for Distributed Agents xt ∈ Ψ. for special classes of cooperative control problems. for all (7. k ≥ t. dierent conditions for . with (7.26) γ1 and γ2 given by (7. On the other hand. Remark 7.1 and 7.24) and γ2 with −1 α4D ◦ ρ−1 ◦ σ4D (7. γ1 where let dene −1 α4 ◦ ρ−1 ◦ σ4 (7. Moreover. the generality of the problem makes it rather dicult to obtain tighter conditions without introducing more restrictive assumptions on the structure of the agents' dynamics and on the cost function. k ≥ t. Now.19) is 0AS in X M P C × IRnz . w)[k] ||)}.23).24) and (7. assume that the following small gain condition holds: γ1 ◦ γ2 (s) < s.3 It is worth noting that the smallgain condition (7.4-7.20)(7.25) ρ any K∞ -function such that (id − ρ) ∈ K∞ .26) may turn out to be conservative in practice as it is typical of these kind of results.8 are veried. the following result about the stability properties of the team of cooperating agents can be proved. Indeed.3 Suppose that Assumptions 7. γ2 (||V (x.25) and argument s takes its val- ues from a suitable subset of IR≥0 according to inequalities (7. for all all and ˆ VD (zk ) ≤ max{βD (VD (zt ).18) and (7. k − t).152 Chapter 7. k ∈ Z≥0 .

the FH cost function (7. team of A M =3 vehicles will be considered. in the special case where the state equation (7. more specialized and tight results can be found.4 As expected. 2004a]. The reader is referred to [Franco et al. by assuming the knowledge of feasibility regions and a specic hierarchical design of the decentralized MPC control problem: the computations are shared by nodes with dierent priorities. As another example. which can impose their control decisions on the subordinate neighbors. 2006] for more details. whose continuous-time models . the control law takes on an explicit feedback feedforward structure and some interesting properties hold.1) takes on a linear structure. In particular.4. Remark 7. Simulation results 153 the stability of the team of agents can be obtained.6) is quadratic. we recall that in [Dunbar & Murray 2006] stability has been shown for formation control of UAV's under dierent hypotheses as the knowledge of the neighbors dynamics. and no state and control constraints are present.4 Simulation results In this section we will show some simulation results concerning a team of UAVs moving in IR2 with nonlinear dynamics. suitably fast information exchange and bounded error between the predicted and actuated state trajectories of each member of the team. 7. stability of a set of decoupled systems is ensured in [Kevicky et al. For instance.7. Such a problem has been selected because of its reasonable simplicity so as to be able to ascertain the basic features and properties of the proposed cooperative control law.

as shown in [Nesic & Teel 2004a]. the simulation trials will refer to the above approximated discrete-time model for mere illustration purposes and to show the eectiveness of the proposed cooperative control scheme. 2. 1999b] for the general case of control of nonlinear sampled-data systems). t. we assume that all the members of the team have the same physical parameters: the mass is inertia is m = 0. θi . The state vector of each agent will be from now on denoted as and is dened by considering the position and velocity in each direction of the plane. This is clearly an important issue (see the above reference for more details and the works [Nesic et al. x ˙ m¨i = −µ1 yi + (uiR + uiL ) sin(θi ). The continuous-time models (7. uiR ) . the linear friction coecient is µ1 = 0. where (7. Nesic et al. For simplicity. However. thus obtaining suitable discretezi t and the the state vectors are denoted by control vectors are denoted by uit . xi .27) are discretized with a sampling time time models. ˙ ˙ whereas the control vector ui col (uiL . yi .27) i = 1.00316 Kgm2 .15 Kg/s µ2 = 0. while the optimization is performed with respect to .9cm.5 In the following. 2004]: m¨i = −µ1 xi + (uiR + uiL ) cos(θi ). 3 .154 Chapter 7.75 Kg . the J = 0. For a MPC algorithm where the continuous time evolution of the system is explicitly taken into account. xi . zi .005 Kgm2 /s and nally the and the rotational friction coecient is radius of the vehicle is rv = 8. where. Cooperative NMPC for Distributed Agents and data are taken according to [Jin et al. Remark 7. yi ) . y ˙ ¨ ˙ J θi = −µ2 θi + (uiR − uiL )rv . 1999a.1s. in some cases the control law that stabilizes the approximated discrete-time model may perform quite poorly when applied to the exact model. plus the orientation angle and rotational velocity is given by zi ˙ col (θi . at time T = 0.

∀j ∈ G1 ). based on leader one. The desired trajectories have been chosen with constant velocities and null rotational velocity. we let zjk = (¯1k − dj1 ) + (zjt−∆ij − z1t−∆ij + dj1 ) .1 with FH cost function t. z1k − di1 ¯ represents the desired trajectory of the dij are such that the three UAVs assume a triangle formation. analogously.4. each agent t+Nip −1 Ji = k=t zik − z1k + di1 ¯ 2 Qi + uik − ui ¯ t+Nip −1 2 Ri + zit+Nip − z1t+Nip + di1 ¯ 2 Pi + k=t j ∈ Gi zik − zjk + dij ˜ 2 Sij (7. . The objective of the distributed cooperative controller is to reach a certain formation following a predened desired trajectory. Specically. For the information vector to take on a constant value within the prediction horizon. Moreover agent 2 gets information from the leader and from agent 3 and.7. . Simulation results 155 a piece-wise constant control signal. ˜ z ¯ The delays have all been set to ∆ij = ∆ = 5T and the communication topology is assumed to be stationary. . see [Magni & Scattolini 2004].2. At every time instant solves a local Problem 7.28) where z1k ¯ represents the desired trajectory of the leader while dij are the desired distance between agent Hence the term The values of The term i and agent j (dii = 0. . for each UAV. ∀i = 1. ui ¯ is the control vector necessary in order to maintain each UAV on the desired trajectory. it is supposed that the leader does not receive any information from the other agents (hence S1j = 0. i − th UAV. agent 3 gets information from the leader and from agent 2. M ). .

i = uiLmin ≤ u1it ≤ 1. For the other agents. 3 .1 · diag (1. 0. + sin(π/3) cos(π/4) − 0.1. from the choice of Qi . are dierent. 3 . S1j = 0. we consider the same values of the parameters. 1) . 2.1) . 1. − sin(π/3) cos(π/4) + 0. though quite small. 0. N1p . comparable since the matrices The control necessary in order to maintain each UAV on the desired trajectory is u1i = 1 . − sin(π/3) cos(π/4) − 0. 0. uiLmax = 6 .1 using The matrices Pi ij . 3 . we have Nic = 10 . i = 2. the terminal constraints zit+Nip − zit+Nip ¯ have been obtained numerically according are constant along the trajectories and These values are not to Lemma 7. 0. 1.1. αwi = 0. by the auxiliary control are obtained. 0) d13 = 16col(0. and The FHOCP is characterized by the constraints with uiRmax . 1) . + sin(π/3) cos(π/4) + 0.2. 3 .01 · diag (1. 1. The values of component of vector uiLmin = 0 . Sij = diag (0. law designed according to Lemma 7.156 Chapter 7. 1.5 cos(π/4). 0. 0.01 · diag (1. Cooperative NMPC for Distributed Agents N1c = N1p = 5 . 1. Ri and S βi = 3 and ˜ Sij = 2Sij . 1. 1. Qi = 0. uiRmin = u1it (u2it ) denotes the rst (second) where Moreover. uiLmin ≤ u2it ≤ uiRmax . 2. G1 . i = 1. u2i = 1 . 3 . 0. 50. uit .1 · diag (1. that is.1. 0. 1) . 0. 0.1. 0) d21 = 16col(0. 1) .5 cos(π/4). 0) . i = 2. ∀j ∈ The values of the parameters used for the leader are Q1 = 0. − sin(π/3) cos(π/4) − 0. + di1 2 i P ≤ Υi . Nip = 250 .5 cos(π/4). uiRmax = 6 . R1 = 0. are respectively Υi Pi Υ1 = 0.96 .5 cos(π/4). The lengths of horizons and N1c .5 cos(π/4). 50. 1. − sin(π/3) cos(π/4) + 0. ¯ ¯ The values of the desired distances between the agents are the following: d12 = 16col(0. Ri = 0. i = 1.3 and Υi = 1.5 cos(π/4). 2. are indeed sucient for the leader to show a reasonably good tracking performance as it starts quite close to the desired trajectory.

ui col(uiL .7. − cos(π/4). It is worth noting the cooperative behavior of the agents: in particular agent 2 (on the left of the leader) even if it starts on its trajectory (z20 = z20 ) ¯ it moves on the right in order to reach faster a better (with Without the cooperative respect to cost (7.5 cos(π/4). z20 = z20 . the dierence between the rst and the second components of the control variables are shown. 0.3(a) and the behaviors of the rst component of the control variables are plotted. This has been done to better appreciate the dierences between the rst and the second . ¯ 1 1 (¯1i + u2i ) cos(π/4). holding constant the velocities. 0) d31 = 16col(0. 0. cos(π/4). 7. uiR ) 7. 0. starting from the initial conditions.2.5 cos(π/4).28)) formation with agent 3. Simulation results 157 d23 = 16col(0.3(c) and 7. the initial condition of the desired trajectory of the leader is: z10 = col (π/4. + sin(π/3) cos(π/4) + 0. 7. the team trajectories are reported in the two-dimensional space: the objective is to attain a triangle formation along a straight line of 45◦ as followers of the leader.4. 0. the behaviors of the control variables of Agents 2 and 3 are shown. 0. 0.3(d). 7. 0. the initial conditions of the UAVs are z10 = z10 . In Fig.3(b) In particular. z30 = z20 + 3. term in the cost function the trajectory of agent 2 would be a straight line.8d23 . The dotted lines depict the actual behavior of the agents. (¯1i + u2i ) sin(π/4)) .3. ¯ ¯ ¯ In Fig. sin(π/3) cos(π/4) − 0. 0. whereas in Figs. 0. 7. 0) . Moreover. − cos(π/4). in Figs. cos(π/4). Finally. 0) and d32 = 16col(0. u ¯ u ¯ m m The entire desired leader's trajectory is obtained.

the problem of designing cooperative control algorithms for a team of distributed agents with nonlinear discrete-time dynamics. 7.158 Chapter 7. the dashed lines depict the constraints imposed on the control variables. components of the control variables. 7. In Fig. actually.5 Conclusions In this chapter.3. these dierences are rather small due to the small magnitude of the variations of the orientation of the two agents.2: Team trajectories (dotted lines). The problem formulation is based on a decentralized MPC framework. The front of the vehicle is represented by the symbol '*' whereas the back of the vehicle is represented by the symbol '+'. where the dynamics of the distributed agents are linked . and analyzing the stability properties of the overall closed-loop system has been investigated. Cooperative NMPC for Distributed Agents Figure 7.

which allows the systematic derivation of rigorous stability results. which is derived by a nonlinear MPC framework.7.5. A key contribution is the general problem formulation. Finally. The resulting local control laws take the form of a feedback feedforward structure. by a cooperative cost function. the team of cooperating agents is treated as a single dynamical system resulting from a feedback interconnection of regionally ISS systems. The stability analysis is made possible by combining the stability properties of the MPC local control laws and ISS arguments. thus allowing the use of small-gain conditions to . which take into consideration delayed state information from neighboring agents. Dashed lines: control constraints. (c) and (d) dierence between the rst and the second components of the control variables. (a) and (b) behaviors of the rst component of the control variables. Each agent uses locally computed control laws.3: Behaviors of the control variables of Agents 2 and 3. Conclusions 159 Figure 7.

wk )) + σ4 (||w||) for all xk ∈ Ω. 2 Moreover (see [Limon et al. wk+1 ∈ W . all wk . w)) where (7. let consider system (7. Future research eorts will be devoted towards (i) considering the case where disturbances and uncertainties aect the communication between the agents of the team and (ii) addressing the robustness issue by generalizing the methodology to the case where optimality of the algorithm is not required at each time-stage. and all k ≥ 0. wk )) + ζ(|wk |) +σ2 (|wk |) + σ3 (|wk+1 |) ≤ −α4 (V (xk . w)). Assume now that xt ∈ Θw . wk ).9). The proof will be carried out in three steps.30) ˜ V (f (xk . wt ) ≤ b(||w||). xk to Then.1: Step 1. Therefore. Cooperative NMPC for Distributed Agents show asymptotic stability. this implies .14) is a RPI it follows that set for system (7. wk+1 ) − V (xk . First. where σ4 (s) = ζ(s) + σ2 (s) + σ3 (s) . there are some important issues requiring further investigation.9) and the state transition from xk+1 : (7. V (x. 2006a]): α3 (|x|) + ζ(|w|) ≥ α3 (|x| + |w|) ≥ α4 (V (x.29) α4 (s) α3 ◦ α−1 (s) 2 is a K∞ -function. it is going to be shown that Θw dened in (7. Then V (xt . Appendix Proof of Theorem 7. w) ≤ α2 (|x| + |w|) and hence |x| + |w| ≥ α−1 (V (x. From the denition of α2 (s) α2 (|x|) + σ1 (|w|) ≤ α2 (|x| + |w|).160 Chapter 7. wk ) ≤ −α4 (V (xk . ∀x ∈ Ω. Despite the general formulation.

5. then ρ ◦ α4 (V (xk . Conclusions ρ ◦ α4 (V (xt .7. hence id(s) > ρ(s). wt ). Firstly. for all Θw is a RPI set for system (7. On the other hand. one has ρ(α3 (|xk |) + ζ(|wk |)) > σ4 (||w||). wk )) > σ4 (||w||). Then ˜ V (f (xt . wt+1 ) ≤ −(id − ρ) ◦ α4 (b(||w||)) + b(||w||) ≤ b(||w||) . α4 ) is a 161 (id − Without loss of generality. wt+1 ) ≤ (id − α4 )(V (xt . Now it is shown that. otherwise take a bigger α2 so that α3 < α2 . wt )) ≤ σ4 (||w||). From the inequality (7. By induction one can show that all ˜ V (f (xt+j−1 . wk+j ) ≤ b(||w||) k ≥ t. ∀s > 0 . it follows that ρ ◦ α4 (b(s)) = σ4 (s) it follows that and. ˜ V (f (xt .29). wt )) + σ4 (||w||) ≤ (id − α4 )(b(||w||)) + σ4 (||w||) = −(id − ρ) ◦ α4 (b(||w||)) + b(||w||) −ρ ◦ α4 (b(||w||)) + σ4 (||w||).9). Step 2. wt ). assume that K∞ -function. (id − ρ) is a K∞ -function. the state tends asymptotically to Θw . wt+j−1 ). if xk ∈ Ω \ Θw . starting from Ξ \ Θw . Hence for j ∈ Z≥0 . owing to (id − ρ) is a K∞ -function. From the denition of the fact that b. that is xk ∈ Θw .

it is true that limk→∞ |x(k. w) ∈ Ω. in turn. wk+1 ) − V (xk . and all wk . w) ∈ Θw . wk+1 ∈ W . w)|Θw = 0. Then ˜ V (f (xk .162 Chapter 7. Moreover. Cooperative NMPC for Distributed Agents then α3 (|xk |) + ζ(||w||) > α3 (|xk |) + ζ(|wk |) > ρ(α3 (|xk |) + ζ(|wk |)) > σ4 (||w||) = ζ(||w||) + σ2 (||w||) + σ3 (||w||) for all xk ∈ Ω \ Θw . the state will reach the region Ω in a nite time. which. x. implies that ˜ V (f (xk . wk ). wk ). and all wk ∈ W . ¯ for all x1 ∈ Ξ \ Ω. Otherwise. wk )) > σ4 (||w||). ¯ Therefore. wk ) ≤ −α3 (|xk |) + σ2 (||w||) + σ3 (||w||) < 0 for all (7. from (7. starting from If in particular Since if Ξ.30) . wk+1 ) − V (xk . ¯ / moreover. x. there exists T1 such that x(T1 . x(T1 .31) it follows that α3 (|x2 |) ≤ α3 (|x1 |) − c. Hence. is achieved in a nite time. there exists x2 ∈ Ω \ Θw ¯ Then from (7. ¯ −α3 (|x1 |) + c ≤ −α3 (|x2 |) < −σ2 (||w||) − σ3 (||w||).14). ¯ x(T1 .31) xk ∈ Ω \ Θw . x. wk+1 ∈ W . w) ∈ Θw . such there exists that c > 0 such that for all x1 ∈ Ξ \ Ω. and all x2 ∈ Ω \ Θw . ρ ◦ α4 (V (xk . x. wk ) ≤ −α3 (|xk |) + σ2 (||w||) + σ3 (||w||) < −¯ c for all xk ∈ Ω \ Θw . the region Θw ¯ Θw is a RPI set. and all wk . in view of (7.

wt ). x. using the same steps of the proof of Lemma 3. γ (||w||)}. It is clear that Let Ψ {x : Ψ V (x. wk ) ≤ max{β(V (xt . all w ∈ MW . γ (||w||)}. k − t). ∀w ∈ W}. Since the upper bound of V (x. ∀w ∈ W}. w)|Θw = 0. where −1 γ = α4 ◦ ρ−1 ◦ σ4 . where the last step is obtained using (7. ∃T2 (ε ) ≥ T1 such that V (xT2 . w) ≤ e = maxR(e)⊆Ω e. wk )) −ρ ◦ α4 (V (xk . ¯ Ξ. wk+1 ) − V (xk . γ (||w||)}.5. ˆ ˆ ˆ ≤ max{β(2α2 (|xt |). Step 3: Given Finally. wk ) ≤ −α4 (V (xk . Therefore. and all wk . ˆ . k − t) + β(2σ1 (|wt |). ˆ Hence ˆ α1 (|xk |) ≤ max{β(α2 (|xt |) + σ1 (|wt |). wk )) ≤ −(id − ρ) ◦ α4 ◦ α1 (|xk |) for all xk ∈ Ω \ Θw . wk )) + σ4 (||w||) = −(id − ρ) ◦ α4 (V (xk . wT2 ) ≤ ε + b(||w||) . wk ).5 in [Jiang & Wang 2001]. that also hold for discontinuous systems. Then. R(e) {x : V (x. w) is known in Ψ⊆Ω then. ∀k ≥ t ˆ for all xt ∈ Υ. wk+1 ∈ W . k − t). the state will arrive close to Θw in a nite time Θw asymptotically. ∀ε > 0.32) Ξ. it is possible to show that there exist a ˆ KL-function β and a K-function γ ˆ such that ˆ V (xk . k − t). ¯ Ψ ⊇ Θw and that is a RPI set. Hence limk→∞ |x(k. it is shown that system (7.9) is regionally ISS in let e ∈ IR≥0 . wk )) + σ4 (||w||) ≤ −(id − ρ) ◦ α4 (V (xk .7. starting from and in (7.11). w) ≤ e. Conclusions 163 one has ˜ V (f (xk .

Using the fact that Υ Ξ is achieved in a nite time. Υ By Lemma 2. by (7. Nic . xt ∈ Υ. ˆ ˆ ˆ β(r + s. . t) are KL-functions while γ1 (s) α1 ◦ˆ (s) and γ2 (s) ˆ ˆ β1 (σ1 (||w||). Then UAG in Ξ. κif (xit+1 ). XiM P C ⊇ Xi κif ⊇ Xif . . then it is ISS in tem admits an ISS-Lyapunov function in Xi In this respect. wit ) o Ji (xit . Proof of Theorem 7. and all w ∈ MW .33) the system (7. where β1 (s. ui[t. t) for all all The last step is obtained (see [Limon et al.1.9) ˆ is ISS in UAG in in Υ. XiM P C . k − t) + β1 (σ1 (|wt |). for any xit ∈ Xif . t) and −1 ˆ ˆ γ ˆ β2 (s. By Theorem 7. k − t) + γ1 (||w||) ˆ ˆ ˆ ≤ β1 (α2 (|xt |). k − t) + γ1 (||w||) ˆ ˆ ≤ β2 (|xt |. γ1 (||w||)} ˆ ˆ ˆ ≤ β1 (α2 (|xt |). k − t).t+Nic −1] .2: the sequence First. Then. k − t) + γ2 (||w||). t) α1 ◦ β1 (2s.7. κif (xit+Nic −1 )] is an admissible control sequence for the FHOCP. Cooperative NMPC for Distributed Agents k ≥ t. 2006a]). ˜ ui[t.33) −1 ˆ xt ∈ Υ. wit . considering also Assumption 7. if sys- M P C . ∀k ≥ t ˆ for all (7. the regional ISS property in Υ is equivalent to and LS. . k − t) + β1 (σ1 (||w||).1. 0) + γ1 (||w||) are K-functions. Nip ) is an ISS-Lyapunov function in XiM P C .4. UAG and LS imply ISS in Υ implies UAG in Ξ. and all w ∈ MW . in the following it will be shown that Vi (xit . considering that Then ˆ ˆ |xt | ≤ max{β1 (α2 (|xt |). t) ≤ β(2r. . .164 Chapter 7. k − t) + β1 (σ1 (|wt |). by Assumption 7. t) + β(2s. t) β1 (α2 (s). So.t+Nic −1] col [κif (xit ).

∀xit ∈ Xi . ∀wit ∈ Wi The lower bound on (7.5-7. Nic . suboptimal) control sequence for the FHOCP . wit ) ≤ k=t (Lli + Lliu Lκif + Lqi )(Lfic )k−t |xit | + Lqiw (αwi )k−t |wit | +βVif (Lfic )Nip |xit |) ≤ (Lli + Lliu Lκif + Lqi ) +Lqiw (Lfic )Nip − 1 |xit | Lfic − 1 (αwi )Nip − 1 |wit | + βVif ((Lfic )Nip |xit |) αwi − 1 Hence there exist two upper bound is veried: K∞ -functions αi2 and σ i1 such that the following Vi (xit .7. Nip ) t+Nip −1 ≤ k=t t+Nip −1 [Lli |xik | + Lliu |κif (xik ) | + Lqi |xik | + Lqiw |wik |] + βVif (|xit+Nip |) (Lli + Lliu Lκif + Lqi )|xik | + Lqiw |wik | + βVif (|xit+Nip |) k=t ≤ so that in view of point 3 of Assumption 7.5: Vi (xit . Conclusions 165 Moreover. in view of Assumption 7. one has t+Nip −1 Vi (xit . wit ) is easily obtained using Assumption 7.4.t+Nic −1] .35) ¯ ui[t+1. wit ) ≤ αi2 (|xit |) + σi1 (|wit |).5). wit ) ≥ ri (|xit |).t+Nic −1] .7 and point 5 of Assumption 7. wit .t+Nic ] col uo i[t+1. it turns out that (7. ∀wit ∈ Wi Now. ui[t. in view of Assumptions 7.5.4 and owing to (7.4 ˜ Vi (xit . ∀xit ∈ Xi κif .34) Vi (xit . wit ) ≤ Ji (xit . κif (xit+Nic ) (7.36) is an admissible (in general.

uo ) − qi (xit . Nic .166 Chapter 7. ui[t+1. ANi −1 wit+1 ) w +Vif (fi (xit+Nip . wit ) + σi2 (|wit |) + σi3 (|wit+1 |) t where σi2 (|wit |) σi3 (|wit+1 |) (αwi )Nip − 1 |wit | + ψi ((αwi )Nip −1 |wit |) αwi − 1 (αwi )Nip − 1 Lqiw |wit+1 | are K∞ -functions. Ak−t wit ) wi wi ≤ qi (xik . wit ) + i k=t+1 [li (xik . uik )+ ¯ k−(t+1) +qi (xik . Ak−(t+1) wit+1 ) − qi (xik . wit+1 .4. Ak−t wit ) wi wi ≤ Lqwi Ak−(t+1) wit+1 − Ak−t wit wi wi = Lqwi (αwi )k−(t+1) wit+1 − Awi wit = Lqwi (αwi )k−(t+1) wit+1 + αwi |wit | and by using point 6 of Assumption 7. κif (xit+Nip ))) − Vif (xit+Nip ) Noting that. wit ) − li (xit . Cooperative NMPC for Distributed Agents t+1 at time with cost ¯ Ji (xit+1 .t+Nic ] . one obtains ¯ Ji (xit+1 . Nip ) ≤ Vi (xit . wit+1 . using Assumption 7.t+Nic ] . Ak−(t+1) wit+1 ) − qi (xik . uok ) − qi (xik . uot ) − qi (xit . Nip ) t+Nip −1 = Vi (xit . κif (xit+Nip )) + qi (xit+Nip . ui[t+1. wit ) + ψi (| (Awi ) t t+Nip −1 Nip −1 wit |) + k=t+1 Lqwi (αwi ) k−(t+1) wit+1 + αwi |wit | ≤ Vi (xit . Nic .6 qi (xik . αwi − 1 αwi Lqiw and . wit ) − li (xit . uo ) − qi (xit . Awi wit+1 ) − li (xik . wit ) − li (xit . Ak−t wit ) i wi ip +li (xit+Nip .

(7. Nic . Owing to the smoothness of Proof of Lemma 7. wit+1 . optimal cost XiM P C Therefore. In order to prove point 6. in view of the admissible is a RPIA set for the closed control sequence (7. wit ) ≤ −ri (|xit |) + σi2 (|wit |) + σi3 (|wit+1 |) for all (7. and point 5 is satised where 3 of Assumption 7. wi ) + ψi (|wi |) (7. and all wit .37) xit ∈ Xi . uo i[t. ∞) such that points 1. Finally.38) . Ki xi ) Πi fi (xi . Nic . and and βVif (| xi |) = λmax (Πi )|xi |2 .1: Ui .8. ui[t+1.8) in is ISS in and hence the closed-loop system XiM P C .4 are satised for with ¯ xi Πi xi ≤ Υi . Conclusions 167 Now. from inequality ¯ Vi (xit+1 . Ki xi ) − xi Πi xi ≤ −xi Qi + Ki Ri Ki xi −qi (xi .5.36). it follows that Vi (xit+1 .t+Nic −1] . Nip ) XiM P C is an ISS-Lyapunov function for the closed-loop system (7. Nip ) and by Assumption 7.34). Ki xi ) − Aicl xi fi (xi . it follows that loop (7. (7. the Ji (xit . wit+1 ∈ Wi . by (7.7. λmin (Πi ) of λmax (Πi ) denote the minimum and the maximum eigenvalues Πi .t+Nic ] .8). wit+1 )−Vi (xit . κif (xi ) and the fact that it is a stabilizing control law and recalling that it follows that there exists 0 ∈ Xi .35). letting Φi (xi ) the inequality fi (xi . wit . wit+1 ) ≤ Ji (xit+1 . respectively. 0 ∈ ¯ Υi ∈ (0.5.37) and Assumption 7. 2. αVif (| xi |) = λmin (Πi ) |xi |2 .

˜ Q it follows ˜ xi Qi xi − xi ˜ ≥ xi Qi xi − xi with Qi + Ki Ri Ki xi − qi (xi . there exists Np such that for all Np ≥ Np .39) Indeed.168 Chapter 7. which implies that inequality (7. ∀xi ∈ Xif . (7. dene chosen (7. wi ) + ψi (|wi |) (7. wi ) + ψi (|wi |) where Lri sup |Φi (xi )| / |xi | . from (7.40) is satised if γi > 0.39) is equivalent to 2Φi (xi ) Πi Aicl xi + Φi (xi ) Πi Φi (xi ) ˜ ≤ xi Qi xi − xi Now. ∀xi ∈ Xif . ∀wi ∈ / . ∀xi ∈ Bri . wi ) + ψi (|wi |) ˜ Qi + Ki Ri Ki xi − xi Si xi ≥ γi |xi |2 . Cooperative NMPC for Distributed Agents is equivalent to 2Φi (xi ) Πi Aicl xi + Φi (xi ) Πi Φi (xi ) + x Aicl Πi Aicl xi − xi Πi xi ≤ −xi Qi + Ki Ri Ki xi − qi (xi .17) it is easy to show that inequality (7. since (7. wi ) and the denition of fi ∈ C 2 ) .41) holds as well.40) Qi + Ki Ri Ki xi − qi (xi . there exists Υi ∈ ¯ 0. Then.41) Lri → 0 as r → 0. xi ∈Bri Bri {xi : |xi | ≤ r} (Once r.38) holds ¯ ¯ Finally. Lri does exist and takes on a nite value because Moreover by the assumption on that qi (xi . γi |xi |2 ≥ {2Lri |Πi ||Aicl | + Lri 2 |Πi |}|xi |2 Hence. Υi such that inequality (7.

thus ending the proof. M where α2 (|xt |) i=1 αi2 (|xt |) and σ1 (|wt |) M i=1 σi1 (|wt |).34) and (7. wt ) ≤ i=1 Clearly αi2 (|xit |) + i=1 i=1 σi1 (|wit |) |xit | ≤ |xt | and |wit | ≤ |wt |. . instead of the above denition of V .18).5. ∀i = 1.35). . κif (xi ))) − Vif (xi ) ≤ −xi Qi + Ki Ri Ki xi 169 −qi (xi .4 is satised. 2 . wi ) + ψi (|wi |) ˜ ˜ < 0 where wi = (Awi )Nip −1 wi ˜ so that point 4 of Assumption 7.7. . wt ) ≤ i=1 M αi2 (|xit |) + i=1 M σi1 (|wit |) σi1 (|wt |) i=1 ≤ i=1 αi2 (|xt |) + ≤ α2 (|xt |) + σ1 (|wt |) . wt ) i=1 for system (7. wit ) 2 From (7. too. a weighted sum of Lyapunov functions could be used along the reasoning provided in [Khalil 2001] in the framework of composite systems. It is worth noting that. it follows that M M M ri (|xit |) ≤ V (xt . Proof of Lemma 7. .2: Consider the ISS-Lyapunov function candidate M V (xt . Conclusions Wi Vfi (fi (xi . M M M and thus V (xt . Vi (xit .

. considering the function ri .37) it follows that (7. . .43) M M ∆V Vi (xit+1 . wt ) ≤ α2 (|xt |) + σ1 (|wt |). ∀xt ∈ X . ∀wt ∈ W V (xt . Cooperative NMPC for Distributed Agents M M M Moreover |xit | ≤ i=1 i=1 Now. . M K are arbitrarily chosen positive scalars). recall that. one has M M M ri (|xt |) ≤ ri i=1 and hence |xit | ≤ i=1 ri (M |xit |) ≤ i=1 ri (M |xt |) ri (|xt |/M ) ≤ ri 1 M M M |xit | i=1 ≤ i=1 ri (|xit |) . . . one (7. . ∀wt ∈ W From (7. M } . wit+1 ) − i=1 M i=1 M Vi (xit . M any K function γ. Therefore. |xt | = M |xt |. for a generic i ∈ {1. for Then |xt | ≥ 1 M |xit | i=1 and M |xt | ≤ i=1 M |xit |.42) r(|xt |) ≤ V (xt .170 Chapter 7. ∀xt ∈ X κf . . wt ). letting has r (|xt |) ri (|xt |/M ) for an arbitrarily chosen index i. one has γ i=1 ai ≤ i=1 γ(M ai ) where ai > 0. Therefore. i = 1. wit ) M ≤ − i=1 ri (|xit |) + i=1 σi2 (|wit |) + i=1 σi3 (|wit+1 |) .

for system (7.43) and (7. wt ) is an ISS-Lyapunov function in X M P C this system is ISS in X MP C.44) xt ∈ X .7.5. Therefore. Let only very briey sketch some parts of the proof just for the purpose of introducing a few quantities that will be used subsequently. by (7. (7. and all w ∈ MW . the proof that it is ISS is obviously trivial since (7. a candidate ISS- .44). Following [Jiang & Wang 2001]. Conclusions 171 Moreover.42). letting σi3 (|wt+1 |) i=1 M σ2 (|wt |) i=1 and σi2 (|wit |) M σ3 (|wt+1 |) i=1 it follows that σi3 (|wit+1 |) ∆V ≤ −r (|xt |) + σ2 (|wt |) + σ3 (|wt+1 |) for all (7.19) is an asymptotically stable discrete-time linear system.19) is concerned (remember that this system describes the eects of the timedelays in the information exchange variables). As far as system (7. note that M − i=1 and ri (|xit |) ≤ −ri (|xt |/M ) M M σi2 (|wit |) ≤ i=1 and σi2 (|wt |) i=1 M M σi3 (|wit+1 |) ≤ i=1 Then.18) and hence V (xt .

0).23) it V (xk . γ1 (βD (VD (zt ). wt ). it is straightforward to obtain (7.172 Chapter 7. Cooperative NMPC for Distributed Agents VD (zt ) Q. 0). γ1 (βD (VD (zt ). 0))} for all xt ∈ Ψ. 0).3: follows that If γ1 ◦ γ2 (s) < s. 0). zt P zt .45) (λmin (P ) and α1D (s) λmin (P )s2 ∆VD and α2D (s) λmax (P )s2 λmax (P ) denote the minimum and maximum eigenvalues of P. γ1 ◦ γ2 (||V (x. wk ) ≤ ||V (x. . γ2 ◦ γ1 (||VD (z)[k] ||)} ˆ ˆ ≤ max{βD (VD (zt ).46) ∆VD ≤ −α3D (|zt |) + σD (|xt |) where α3D (s) 1 2 2 λmin (Q)s and σD (s) P AA λmax ( 2B λmin (Q)P B + B P B)s2 . Moreover.13 in [Jiang & Wang 2001]. wk ). respectively). wt ). and all k ∈ Z≥0 .19) is where is the for a positive denite solution of the Lyapunov equation given symmetric positive-denite matrix that A P A − P = −Q It can be immediately shown α1D (|zt |) ≤ VD (zt ) ≤ α2D (|zt |) where (7. P Lyapunov function for system (7. from (7. w)[k] || ˆ ˆ ≤ max{β(V (xt . γ2 (β(V (xt . Proof of Theorem 7. 0)). dening VD (zt+1 )−VD (zt ). 0).22) and (7. k ≥ t and hence V (xk . w)[k] ||)} ˆ ˆ ≤ max{β(V (xt . k ≥ t and VD (zk ) ≤ ||VD (z)[k] || ˆ ˆ ≤ max{βD (VD (zt ). 0)} for all xt ∈ Ψ. wt ). VD (zk ) are bounded by initial condition. γ2 (β(V (xt . and all k ∈ Z≥0 . wt ). By Lemma 3.

wk ) ≤ = k→∞ k→∞ −1 lim [α4 ◦ ρ−1 ◦ σ4 (VD (zk ))] lim γ1 (VD (zk )) and k→∞ lim VD (zk ) ≤ = k→∞ k→∞ −1 lim [α4D ◦ ρ−1 ◦ σ4D (V (xk . the system is 0-AS in X M P C × IRnz . wk ) γ1 173 an asymptotic gain from asymptotic gain from to is given by whereas an V (xk . Conclusions VD (zk ) V (xk . the assumption that γ1 ◦ γ2 (s) < s k→∞ lim V (xk . wk ) ≤ ≤ k→∞ k→∞ lim γ1 (VD (zk )) lim γ1 ◦ γ2 (V (xk . wk ) = lim VD (zk ) = 0 k→∞ Thus.47) Again.5. wk ) to VD (zk ) is given by γ2 . wk )) Hence k→∞ lim V (xk .7. Hence: k→∞ lim V (xk . wk )) implies that (7. . wk ))] lim γ2 (V (xk .

.

Beta cells normally produce insulin. variant. Bi- ological systems are highly non-linear and there is a large inter-individual variability. together with glucagon (an other hormone produced by alpha cells of the pancreas). levels. the human body is a system intrinsically time- . occurs when the body's immune system attacks and destroys beta cells of the pancreas. or blood sugar. a hormone that regulates. also called juvenile diabetes or insulin-dependent diabetes. A typical medical application is glycemic control for the type 1 (insulindependent) diabetic patient. There are two major types of diabetes: type 1 diabetes and type 2 diabetes. although the progress has been slow in some cases due to particular challenges encountered by the inherent complexity of biological systems. Type 1 diabetes. the blood glucose Furthermore.1 Introduction Control systems theory has been extending into many elds and medicine is not an exception (see for example [Palerm 2003] and the references therein).Chapter 8 MPC of Glycaemia in Type 1 Diabetic Patients 8. Diabetes (medically known as diabetes mellitus) is the name given to disorders in which the body has trouble regulating its blood glucose.

glucagon increases it. including blindness and kidney failure in the long term. When the beta cells are destroyed. an articial pancreas can potentially result in: 1) less glycemic variability. while insulin lowers it. No articial pancreas system is currently approved. Compared to currently applied intensive therapy. which for control of blood glucose in patients with diabetes is known as an articial pancreas. where it can cause serious damage to all the organ systems of the body. merely an open-loop control with intermittently monitored glucose levels resulting in intermittently administered insulin doses that are manually administered by a patient using a written algorithm. devices that could become components of such a system are now becoming commercially available [Klono 2007]. 3) less pain from pricking the skin to check the blood glucose and deliver insulin boluses. 2) less hypoglycemia. which results in continuously infused insulin dosed according to a computerized algorithm without a need for patient input. MPC of Glycaemia in Type 1 Diabetic Patients concentration in the body.e. no insulin can be produced. and loss of consciousness and coma in the short term when hypoglycemic. or having insulin delivered through an insulin pump. people with type 1 diabetes must take insulin in order to stay alive. This means undergoing multiple injections daily (normally coinciding with the meal times). however. In this case. glucose levels are monitored continuously. This method is approximate. The currently available intensive therapy can be contrasted to glucose management by a closed-loop system.176 Chapter 8. The food intake in the body results in an increase in the glucose concentration. and 4) less overall patient eort. and the glucose stays in the blood instead. Insulin permits most of the body's cells to take up glucose from the blood and use it as energy. Poor glycemic control leads to severe health complications. and testing their blood sugar by pricking their ngers for blood six or more times a day [JDRF Website ]. For this reason. . i.

idiopathic). Closed-loop control for regulating glycemia requires three components: glucose measurements. This number is likely to more than double by 2030 [WHO Website ]. such.8. some are quickly absorbed. These cases are those attributable to an autoimmune process. People with this type of Diabetes Mellitus fully depend on exogenous insulin. The administration of insulin.2 Articial Pancreas Type 1 diabetes (T1DM) is caused by an absolute deciency of insulin secretion. It includes cases primarily due to β cell destruction. the rest of this section covers the various options. It is this specic group which would ben- et the most from closed-loop glucose regulation. a substantial gure to warrant relevant research in this area. in order to regulate fasting levels. and who are prone to ketoacidosis. the control algorithm and a way to administer the insulin (see [Palerm 2003] for details).2. The type of insulin used also varies. as there is the regular human insulin plus other variations that have been engineered to have specic absorption properties.e. or to quickly bring insulin levels up to coincide with a meal. Research towards the realization of an articial . is someAs thing that is relevant even with the current modalities of treatment. The two main modalities currently in clinical use are multiple daily injections (MDI) and continuous subcutaneous insulin infusion (CSII) with an externally worn infusion pump that operates in open-loop. Glucose sensing is dealt in the modeling section below. and type of insulin. Articial Pancreas 177 The World Health Organization (WHO) estimates that in 2006 more than 180 million people worldwide have diabetes. 8. as well as those with β cell destruction for which no pathogenesis is known (i. This allows for a single shot to have an eect over several hours. while others are slow acting.

there is a large inter-individual and intra-individual variability in diabetic patients. 2004]. without counter-regulation: that increases the risk of hypoglycaemia. It adopted the intravenous (iv) route both for glucose monitoring and insulin infusion: an early version used a glucose infusion to counter-regulate insulin action and prevent hypoglycaemia. PID (Proportional-Integral-Derivative) strategy is certainly one of the most simple and follow. 2006] use a PID switching strategy at meals with an algorithm that generates a time-variant setpoint (reference governor ). mimicking β cell response observed in in-vitro experiments. [Marchetti et al.178 Chapter 8. however. These technological improvements have made more feasible the adoption of a subcutaneous route (sc-sc). calculates current insulin infusion as a sigmoidal function of actual blood glucose level. based on Albisser et al. 2003]) and less invasive glucose sensors have been introduced. The iv-iv approach However. there are still diculties that justies the development of specic and sophisticated control algorithms. which are based on dierent approaches. [Albisser et al. partly as a result of studies that show the similarity with the β cell functionality [Steil et al. The control algorithm. control can only act on insulin infusion. Lispro insulin [Homko et al. Some . Furthermore. 2006] have proposed a controller based on PID with time-variant parameters to manage the dierent response between day and night. this is is highly invasive. There is also the need of controller personalization: in fact. [Steil et al. Recently new fast insulin analogues (e. Marchetti et al. Main issues are delays and the presence of important disturbances on glucose level like meals and physical activity. the best approach in terms of performance because it minimizes the delay between insulin infusion and action. Many solutions have been proposed in literature for this control problem. MPC of Glycaemia in Type 1 Diabetic Patients pancreas is active since 1970.g. Steil et al. so it can't be used on a daily basis. The rst prototype commercialized was the so-called Biostator . 1974].

A pure feedback control. which had already been applied successfully in other biomedical issues such as regulation of blood pressure [Gopinath et al. can obtain benets comparable to those using regular insulin via the intravenous route. based on a model of the system identied by the step response. [Shimoda et al. Articial Pancreas 179 authors have shown that integral action is fully responsible for the administration of excessive insulin in the postprandial period. therefore. Parker et al. such as PID. whose parameters are adjusted over time using a Bayesian estimate. The main benet of this strategy is to predict the evolution of blood glucose and. the presence of delays results in unsatisfactory performance. 1999] use a linear predictive control. This technique has the disadvantage compared to linear MPC.8. This has led to the study of solutions based on more sophisticated approaches. Shimoda et al. is able to react against a disturbance only when it is manifested as a variation of the measured variable. with the result of generating numerous hypoglycaemia. not to lead to a closed form for the control law: so the optimization problem must be solved . [Parker et al. known as Lispro. 1995] and of anesthesia [Linkens & Mahfouf 1994]. even if only partially. [Hovorka et al. 2004] have proposed a regulator that uses a nonlinear and time-variant model for the prediction.2. Hovorka et al. to be able to act on time by avoiding hypoglycaemic and hyperglycaemic events. This ability is very useful especially if the disturbances trend is known in advance. The predictive control uses a model to predict the dynamic of the system within a certain time and calculates the input through the optimization of a tness function in that time interval. such as predictive control or MPC. For example. 1997] have used an algorithm PD (Proportional-Derivative) in tests on diabetics. This work has also shown that the use of a rapid insulin analogue. Another advantage is to explicitly consider the presence of constraints on the inputs and outputs. the simplest of that class. In the case sc-sc.

8.S. 2006]). The objective of the project is to make available on the market within 5 years (i. Yale. The algorithms are tested on virtual patient populations and the performances are evaluated by using particular metrics described later on. . the largest nonprot foundation in the world who work for type 1 diabetics. regulatory institutions (FDA.[Dua et al. by 2011) articial pancreas technology and.c. the University of Colorado. the feedback control of glucose concentration in Type1 diabetic patients using s. linear and nonlinear MPC controller synthesized on a recently developed nonlinear in-silico model of glucose metabolism (see [Magni et al. Another approach that allows to solve the problem of constrained optimization o-line needs the use of a parametric programming algorithm (Dua et al. subsequently. To achieve these goals. The University of Pavia collaborates in this project joined with the University of Virginia and the University of Padova (see Figure 8.180 Chapter 8.c. has launched a project to accelerate the development and adoption of a device for the closed loop control of blood glucose. 2007b]) are proposed.2. MPC of Glycaemia in Type 1 Diabetic Patients on-line and can be very computationally expensive. the Sansum research center and the University of Virginia. insulin delivery and s. In particular.1 Articial Pancreas Project In 2006.1). the Juvenile Diabetes Research Foundation (JDRF [JDRF Website ]). launched a campaign to expedite the approval of this device by the U.e. Stanford. JDRF has created a consortium (the Articial Pancreas Consortium ) of British and Americans research centers that currently includes the University of Cambridge. Food and Drug Administration) and to obtain coverage by health insurance. continuous glucose monitoring is considered. JDRF has to promote its use on a large scale. In this chapter.

as substitutes of animal tests.8. experiments using the developed linear model predictive controller are in progress at the Charlottesville and Padova hospitals. Three categories need to be considered: the gas- trointestinal (GI) system as related to digestion and nutrient absorption.1: JDRF Consortium The Food and Drug Administration (FDA) accepted the in-silico experiments. Currently. based on a metabolic simulator and virtual patient populations. the hepatic function in metabolism and the endocrine pancreas with its se- .3 Glucose Regulation There are several topics in physiology that are related to an understanding of glucose regulation and the particulars that need to be considered in modeling and control. Glucose Regulation 181 Figure 8. 8.3.

182 Chapter 8. MPC of Glycaemia in Type 1 Diabetic Patients

cretion of insulin and glucagon as the main regulators of blood glucose level (see [Palerm 2003] for details). The root source of glucose for the body is the food. As ingested, food cannot be absorbed; it thus requires both mechanical and chemical digestion. It is the job of the GI system to do this processing, and then to

facilitate the absorption of nutrients. The GI tract is composed of hollow organs connected in series; these have sphincters at key locations in order to regulate ow. Associated with the GI system are organs and glands that secrete various substances into the hollow organs. Carbohydrates, are present in food as disaccharides and monosaccharides (including glucose); since only monosaccharides are absorbed, carbohydrates require chemical digestion as well. Digestion duration varies with the composition of the meal, but usually it lasts about two hours. At end glucose is absorbed in the intestine, entering bloodstream. In carbohydrate metabolism and the regulation of blood glucose levels the liver plays a central role. It receives blood from the portal circulation (coming from the stomach, small and large intestines, pancreas and spleen). Thus most nutrients (lipids are absorbed into the lymphatic circulation) and pancreatic hormones pass through the liver before continuing on. The liver stores several substances, including carbohydrates. Depending on metabolic requirements these are either sequestered or released into the blood stream. Among these is glucose, which it can also synthesize. Given its central position, the liver plays a critical role in energy metabolism. The two hormones with the largest role in glucose regulation are insulin and glucagon. Between meals, insulin levels are low and glucagon levels are high; given these conditions, the liver serves as a source of glucose. It does this by gluconeogenesis ( synthesis of glucose from amino acids and lactate)

8.3. Glucose Regulation

183

α cells (glucagon)

β cells (insulin) δ cells (somatostatin)

F cells

Blood ows from the center to the perifery

Islet of Langerhans

Figure 8.2: Structure and functionality of islet of Langerhans

and glycogenolysis (the breakdown of glycogen stores into glucose). Other monosaccharides are largely converted into glucose. In the postprandial

period, insulin levels are higher. The liver then takes up glucose from the portal circulation; it either stores it by synthesizing glycogen, or breaks it down to pyruvate (glycolysis). Carbohydrates not oxidized nor stored as

glycogen are metabolized to fat. Insulin and glucagon are secreted by the pancreas, in the islets of Langerhans, the endocrine glands. Each islet contains four types of cells (see Figure 8.2).

β

cells (the most

numerous) secrete insulin, proinsulin, C peptide and amylin. crete glucagon,

α

cells se-

δ

cells secrete somatostatin and F cells secrete pancreatic

polypeptide. These cells are inuenced by external links (neural, cell-cell and humoral), as well as by other cells in the islet. Insulin's actions serve to replenish fuel reserves in muscle, liver and adipose tissue. During fasting,

β

cells secrete small amounts of insulin. With low insulin levels, lipids are

184 Chapter 8. MPC of Glycaemia in Type 1 Diabetic Patients

mobilized from adipose tissue, and amino acids are mobilized from protein stores in muscle and other tissues. These lipids and amino acids are precursors for hepatic ketogenesys and gluconeogenesis, and are also oxidized as fuel. Insulin secretion increases with feeding. Elevated levels of insulin reduce lipid and amino acid mobilization within fuel stores, which are also replenished by the enhanced uptake of carbohydrates, lipids and amino acids by insulin-sensitive target tissues. This action maintains plasma glucose within tight limits, assuring a constant supply of fuel for the central nervous system (CNS). Plasma glucose below 2-3 mM/l (36-54 mg/dl), hypoglycemia, results in confusion and seizures, and if not resolved, coma and eventually death. Severe hyperglycemia (plasma glucose above 30-40 mM) results in osmotic diuresis, leading to severe dehydration, hypertension and vascular collapse. Insulin synthesis and secretion in the exposure to glucose, which is its main regulator. Glucagon promotes hepatic glucose production; it does so by activating glycogenolysis and inhibiting glycogen synthesis and glycolysis. In the liver, glucagon stimulates fat oxidation as a source of energy. If the needs for

β

cells is stimulated by

energy of the liver are exceeded, fatty acids are only partially oxidized, forming ketone bodies. These keto acids can then be used by other tissues as fuel (besides glucose, keto acids are the only other energy source the brain will utilize). This is particularly important during fasting. Other

hormones have minor eects, for example, somatostatin, secreted by

δ

cells

in the pancreas and the hypothalamus, inhibits the secretion of several other hormones, including insulin and glucagon. It is still unclear if it has any paracrine eects on

α

or

β

cells.

8.4. Models for the Glucose-Insulin System

185

**8.4 Models for the Glucose-Insulin System
**

Mathematical models for the glucose-insulin system could be very useful for the research about automatic glycemic control. In fact, they permit the design and validation of control algorithms reducing the number of in-vivo experiments required. Dierent in models from have simple via the been linear proposed models in like literature Ackerman of ranging

complexity

et

al.

[Gatewood et al. 1968],

classical

minimal

model

Bergmann

et al.

[Bergman & Urquhart 1971], to the non-linear models and inclu[Hovorka et al. 2004] and Sorensen et al.

sive models of Hovorka et al. [Sorensen 1985].

Recently, however, thanks to new types of OGTT

1 or Mixed Meal ex2 and

periments using radioactive tracers and new techniques such as NMR

PET , quantitative data have been obtained on physiological variables never seen before, such as the liver production or glucose muscular utilization. Thanks to this new database was possible to develop new models more complex and detailed than previous: in particular, the model developed by Dalla Man et al. (Italy). [Dalla Man et al. 2007b] from the University of Padova

3

8.4.1 Dalla Man et

al.

model for the diabetic patient

The development of this model was made possible thanks to a database of 204 healthy subjects, who was given a mixed meal containing glucose labeled with a radioactive tracer. At the same time, two tracer are ad-

1 2

Oral Glucose Tolerance Test. Nuclear Magnetic Resonance. 3 Positron Emission Tomography.

186 Chapter 8. MPC of Glycaemia in Type 1 Diabetic Patients

ministered intravenously: this triple-tracer technique [Basu et al. 2003] has enabled the ow measurement of glucose in the body. Without the use

of this technique it is only possible to measure blood glucose levels and insulinemia, without being able to distinguish, for example, between glucose produced endogenously by the liver and glucose introduced with the meal. So it was possible to obtain direct measurements of the rate of appearance

Ra,

namely the glucose ow entering the bloodstream from the

gastro-intestinal tract, endogenous glucose production tion

EGP ,

renal extrac-

E

and glucose utilization

U.

These new informations have enabled a

qualitative leap in modeling, with the possibility of splitting the system into sub-units interconnected: for each unit is postulated a compartment model whose parameters are estimated using the available measures of inputs and outputs. The model described in [Dalla Man et al. 2007b] refers to an healthy person. Therefore, it was necessary to make some structural changes to the model to reect the absence in patients with type 1 diabetes of endogenous insulin production (see [Magni et al. 2007b]). The fundamental change concerns the replacement of the subsystem of the pancreatic

β -cell

with a unit

that models the dynamics of a subcutaneous (sc) infusion of insulin. The model, overall, may be seen as a MISO (Multi Input Single Output) with two inputs, the glucose introduced with the meal and insulin administered subcutaneously, and only one exit, the blood glucose. In the following sections the functional units are described in details. interconnections between the units, see Figure 8.3. For an overview of

Models for the Glucose-Insulin System 187 Insulin Meal S.3: Interconnections between subsystems of Dalla Man et al. 2007b] changed to cope with type 1 diabetic patients as described in [Magni et al.4: Schema of gastro-intestinal subsystem .8.C. Glucose subsystem Glucose subsystem Renal Extraction Insulin-independent utilization Degradation Utilization subsystem Insulin-dependent utilization Figure 8. Insulin subsystem EGP subsystem Gastro-Intestinal subsystem Absorption EGP Rate of Appearance Insulin subsystem S.Qsto) Qgut kabs Ra(t) Figure 8.4.C. 2007b] d(t) Qsto1 kgri Qsto2 kempt(t. model [Dalla Man et al.

The glucose ow d(t) (mg/min) coming from the meal is the input of the rst kgri . compartment.4): two for the glucose in the stomach (solid liquid Qsto1 and Qsto2 (mg)) and one for the glucose in the intestinal tract Qgut (mg). f is considered constant (f = 0. Furthermore. Qsto1 (t) = −kgri Qsto1 (t) + d(t) ˙ ˙ Q (t) = −k sto2 empt (t. (8. Qsto (t)) · Qsto2 (t) + kgri Qsto1 (t) ˙ Q (t) = −kabs Qgut (t) + kempt (t.1) kgri is xed and equal to kmax . The stomach digests the meal with grinding coecient then the chyme enters the intestine with fractional coecient of transfer kempt . Qsto (t)) = kmax + . that is a time-variant nonlinear function of total Qsto = Qsto1 + Qsto2 that will be later described. Qsto (t)) · Qsto2 (t) gut Qsto (t) = Qsto1 (t) + Qsto2 (t) Ra(t) = f · k · Q (t)/BW abs gut In order to guarantee model identiability.188 Chapter 8. It consists of three compart- ments (see Figure 8. nally glucose is absorbed and enters the bloodstream. The rate of appearance Ra (mg/Kg/min) is a constant percentage f (about 90%) of the total ow leaving the intestine. Coecient of gastric emptying kempt The coecient of gastric emptying linear function of kempt (1/min) is a time-variant non- Qsto kmax − kmin {tanh[α(Qsto (t) − b · D(t))] 2 − tanh[β(Qsto (t) − d · D(t))]} kempt (t. MPC of Glycaemia in Type 1 Diabetic Patients Gastro-Intestinal subsystem Gastro-Intestinal subsystem describes the glucose transfer coming from a meal through the gastro-intestinal system.9).

see [Dalla Man et al. model parameters (see Figure 8.5: kempt (t. Glucose ows (mg/Kg/min) entering the rst compartment are EGP coming from the liver and Ra coming from the gastro-intestinal tract. Qsto ) function. 2006]. d. 2D(t)(1 − b) tf β= d(t)dt 5 2D(t)d D(t) = ti with ti and tf respectively start time and end time of the last meal. b. where D is the total glucose quantity of the last meal where α= 5 .5). Glucose subsystem Glucose subsystem model consists of two compartments (see Figure 8. kmax and kmin For details.6).8. Uid and independent As outputs there are insulin-dependent utilization Uii .4. one for plasma glucose Gp (mg/Kg) and one for the glucose on tissue Gt (mg/Kg). Models for the Glucose-Insulin System kempt 189 kmax kmean kmin 0 bD dD D Qsto Figure 8.

(8. Subsystem equations are Gp (t) = −k1 Gp (t) + k2 Gt (t) + EGP (t) + Ra(t) − Uii − E(t) ˙ ˙ G (t) = k1 Gp (t) − k2 Gt − Uid (t) t G(t) = Gp (t)/VG where (8.190 Chapter 8. in this model is considered constant and equal to 1 mg/Kg/min.3) Rab = 0. MPC of Glycaemia in Type 1 Diabetic Patients EGP(t) k1 Ra(t) Gp k2 E(t) Uii(t) Gt Uid(t) Figure 8. noting that at basal equilibrium (8. i.2) VG (dl/Kg) is the distribution volume of glucose.4) EGPb = Uii + Uidb + Eb Renal Extraction E . is characterized − k1 Gpb + k2 Gtb + EGPb + Rab − Uii − Eb = 0 k G − k G − U = 0 1 pb 2 tb idb so. G (mg/dl) is the glycemia. that occurs primarily in cen- tral nervous system. constant glycemia by the following equations Gb (mg/dl). Basal steady state.e.6: Schema of glucose subsystem and the renal extraction E. Insulin-independent utilization Uii .

8. Models for the Glucose-Insulin System 191 Renal extraction represents the glucose ow which is eliminated by the kidney. coming from the following dynamic system ˙ I1 (t) = ki I(t) − ki I1 (t) I (t) = k I (t) − k I (t) ˙d i 1 i d where (8. when glycemia exceeds a certain threshold ke2 E(t) = max(0.5) Endogenous glucose production EGP EGP comes from the liver. EGPb − kp2 (Gp (t) − Gpb ) − kp3 (Id (t) − Ib )) where (8. where a glucose reserve exists (glycogen). ke1 · (Gpb − ke2 )) Basal renal extraction is null in almost all patients. EGP is inhibited by high levels of glucose and insulin EGP (t) = max(0. ke1 · (Gp (t) − ke2 )) The parameter ke1 (1/min) represents renal glomerular ltration rate.7) I (pmol/l) is plasma insulin concentration or insulinemia and ki (1/min) is a model parameter.8) . At basal one has Eb = max(0.6) kp2 and kp3 are model parameters and Id (pmol/l) is a delayed insulin signal. Insulin-dependent utilization Uid It depends non-linearly from tissue glucose Uid (t) = Vm (X(t)) Gt (t) Km + Gt (t) (8. (8.4.

enters the bloodstream and is degradated in the liver and in the periphery Ip (t) = −(m2 + m4 )Ip (t) + m1 Il (t) + s(t) ˙ ˙ I (t) = −(m1 + m3 )Il (t) + m2 Ip (t) l I(t) = Ip (t)/VI where (8. .12) VI (l/Kg) is the distribution volume of insulin.11) Insulin subsystem Insulin ow s. m3 e m4 (1/min) are model parameters. m2 . coming from the subcutaneous compartments.10) Gtb = (Uii − EGPb + k1 Gpb + Eb )/k2 Vm0 = (EGPb − Uii − Eb ) · (Km + Gtb)/Gtb (8. MPC of Glycaemia in Type 1 Diabetic Patients Vm X where (1/min) is a linear function of interstitial uid insulin (pmol/l) Vm (X(t)) = Vm0 + Vmx X(t) which depends from insulinemia in the following way ˙ X(t) = −p2U X(t) + p2U (I(t) − Ib ) where (8.192 Chapter 8.9) p2U (1/min) is called rate of insulin action on peripheral glucose. Considering basal steady state Uidb = Vm0 and Gtb Km + Gtb (8. m1 .

while CLins (l/min) is the insulin clearance. m4 m1 193 depend on in the following way CLins HEb · VI · BW HEb m3 = m1 · 1 − HEb CLins m4 = 0.15) s = (m + m )I − m I 2 4 pb 1 lb b u(t) S1 ka2 m1 kd S2 ka1 m3 Il m2 Ip m4 Figure 8.4 · VI · BW m2 = 0. At basal one has 0 = −(m2 + m4 )Ipb + m1 Ilb + sb 0 = −(m + m )I + m I 1 3 lb 2 pb so (8.6.8.7: Schema of insulin subsystem .6 · where (8. Models for the Glucose-Insulin System m2 .14) I = lb m2 Ipb m1 + m3 (8. HEb is considered constant and equal to 0.13) HEb (adimensional) is the basal hepatic insulin extraction.4. m3 .

17) sb S1b = kd + ka1 kd S2b = S1b ka1 u =s b b (8.C. This delay is modeled here with a two compartments.18) The quantity ub (pmol/min) represents insulin infusion to maintain diabetic patient at basal steady state. which represent respectively polymeric and monomeric insulin in the subcutaneous tissue S1 (t) = −(ka1 + kd )S1 (t) + u(t) ˙ ˙ S (t) = kd S1 (t) − ka2 S2 (t) 2 s(t) = ka1 S1 (t) + ka2 S2 (t) where (8. . unlike in the healthy subject in which the pancreas secretes directly into the portal vein. Insulin subsystem Usually in diabetic patients. At basal one has ka1 and ka2 are absorption constants. MPC of Glycaemia in Type 1 Diabetic Patients Ipb = Ib · VI . where S.16) u(t) (pmol/Kg/min) represents injected insulin ow. Insulin takes some time to reach the circulatory apparatus. kd is called degradation constant. S1 and S2 (pmol/Kg). insulin is administered by subcutaneous injection. 0 = −(ka1 + kd )S1b + ub 0 = kd S1b − ka2 S2b sb = ka1 S1b + ka2 S2b and solving the system (8.194 Chapter 8.

[Dalla Man et al. .3 respectively. This dynamic was modeled with a system of the rst order ˙ GM (t) = −ksc GM (t) + ksc G(t) (8. Glucose subsystem Subcutaneous glucose GM (mg/dl) is.c.4.1 with basal values. highly correlated with plasma glucose.19) Model states and parameters Final model consists of 13 states. glucose Basal Value 0 0 0 Gpb Gtb Ib Ib 0 Ilb Ipb S1b S2b Gb Table 8. monomeric insulin S. at steady state. Dalla Man et al.8. it follows the changes in plasma glucose with some delay.1: State table for Dalla Man et al. dynamically.c. polymeric insulin S. and 5 constants reported in Table 8. 2007b] model is uniquely dened by 25 independent parameters.C. instead.2 and Table 8. reported in Table 8. Models for the Glucose-Insulin System Stato x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11 x12 x13 Qsto1 Qsto2 Qgut Gp Gt I1 Id X Il Ip S1 S2 GM 195 Unit mg mg mg mg/Kg mg/Kg pmol/l pmol/l pmol/l pmol/Kg pmol/Kg pmol/Kg pmol/Kg mg/dl Description Solid glucose in stomach Liquid glucose in stomach Glucose in intestine Plasma glucose Glucose in tissue Delayed insulin signal Delayed insulin signal Interstitial uid insulin Liver insulin Plasma insulin S. model S.c.

196 Chapter 8.3: Constant table for Dalla Man et al. model . Constant f Uii HEb ke1 ke2 Unit adimensional mg/Kg/min adimensional min−1 mg/Kg Table 8. MPC of Glycaemia in Type 1 Diabetic Patients Parameter BW kabs kmax kmin b d ka1 ka2 kd m1 CLins VI k1 k2 VG EGPb kp2 kp3 ki Ib Km Vmx p2U ksc Gb Unit Kg min−1 min−1 min−1 adimensional adimensional min−1 min−1 min−1 min−1 l/min l/Kg min−1 min−1 dl/Kg mg/Kg/min min−1 min−1 min−1 pmol/l mg/Kg min−1 min−1 min−1 mg/dl Table 8. model.2: Parameter table for Dalla Man et al.

b and g are xed constant. to evaluate the clinical risk related to a particular glycemic value BGI(·) = 10(g[lna (·) − b])2 where (8.5. because hypoglycemia is considered . but also must avoid even isolated episodes of hypoglycemia.e. the value whose risk is zero.07 g = 0. Blood Glucose Index (BGI) Blood Glucose Index is a metric proposed by Kovatchev et al.44 The target b = 10.75 G0 . which is to maintain glucose levels between 80 and 120 mg/dl in the face of disturbances such as meals or physical activity.8. A metric has to take into account these features: other metrics such as average blood sugar or measures related to it as HbA1c are not very signicant. 2005].31 mg/dl The BGI function is asymmetric. A good algorithm should be able to maintain blood sugar low enough.5 Metrics to asses control performance In evaluating the performance of a control algorithm one has to remember that its basic function is to mimic as best as possible the feature of the β -cells. i. Metrics to asses control performance 197 8. as this reduces the longterm complications related to diabetes. [Kovatchev et al. In particular. these parameters are equal to a = 1.20) a. is G0 = e √ a b = 148.

MPC of Glycaemia in Type 1 Diabetic Patients 100 LBGI 90 80 HBGI 70 60 BGI 50 40 30 20 10 0 0 50 100 150 200 250 300 350 400 450 500 550 600 Plasma Glucose (mg/dl) Figure 8. . two synthetic indices for sequences of surements of blood glucose can be dened n mea- LBGI (Low BGI) measures hypoglycemic risk LBGI = 10 n n rl2 (Gi ) i=1 where rl(·) = min(0.198 Chapter 8. g[lna (·) − lna (G0 )]). g[lna (·) − lna (G0 )]) HBGI (High BGI) measures hyperglycemic risk HBGI = 10 n n rh2 (Gi ) i=1 where rh(·) = max(0.8: Blood Glucose Index more dangerous than hyperglycemia (see Figure 8.8). Starting from the BGI.

The CVGA provides a simultaneous assessment of the quality of glycemic regulation in all patients. As such it may play an important role in the tuning of closed-loop glucose control algorithms and also in the comparison of their performances.9: Control Variability Grid Analysis with summary Control Variability Grid Analysis (CVGA) The Control Variability Grid Analysis (CVGA) is a graphical representation of min/max glucose values in a population of patients either real or virtual. Note that .5. Assuming that for each subject a time series of measured Blood Glucose (BG) values over a specied time period (e.8. For each subject a point is plotted whose X-coordinate is the minimum BG and whose Y-coordinate is the maximum BG within the considered time period (see Figure 8. 1 day) is available.g. the CVGA is obtained as follows. Metrics to asses control performance 199 Summary: A=8% B=76% C=4% D=8% E=4% >400 Zone C high Maximum Glucose (mg/dl) Zone D high Zone E 300 Zone B high 180 Zone B Zone D low Zone A 110 90 Zone B low 70 Minimum Glucose (mg/dl) Zone C low <50 Figure 8.9).

70 < Gmin < 90) poglycemia. 9 rectangular zones are dened as follows Zone A (Gmax < 180. Gmin < 70) : Erroneous control. Gmin > 90) : Accurate control. Gmin < 70) poglycemia. 70 < Gmin < 90) perglycemia. overCorrection of hypo- Zone C low (Gmax < 180. : failure to Deal with hy- Zone D low (180 < Gmax < 300.200 Chapter 8. MPC of Glycaemia in Type 1 Diabetic Patients the X axis is reversed as it goes form 110 (left) to 50 (right) so that optimal regulation is located in lower left corner. 70 < Gmin < 90) Zone C high (Gmax > 300. Zone B high (180 < Gmax < 300. Dierent regions on the plane can be associated with dierent qualities of glycemic regulation. : : Benign control. : Benign trend towards Zone B low (Gmax < 180. : failure to Deal with hy- Zone E (Gmax > 300. Gmin > 90) glycemia. Gmin > 90) hyperglycemia. In order to classify subjects into categories. . A synthetic numeric assessment of the global level of glucose regulation in the population is given by the Summary of the CVGA. Gmin < 70) glycemia : overCorrection of hyper- Zone D high (Gmax > 300. : Benign trend towards hy- Zone B (180 < Gmax < 300. The appearance of the overall plot is a cloud of points located in various regions of the plane.

1 Unconstrained LMPC The LMPC regulator design needs to identify a linear-low order model which approximates the behavior of the nonlinear system. In the unconstrained case. xed in in order to obtain a time-invariant model.8.6. It will be used as the prediction model of the LMPC controller. the output variable.e. i.21) x(t) ∈ R13 is the state vector of Dalla Man et al. u(t). Qsto ) = kmean = kmin +kmax . LMPC gives an explicit control law. u(t) ∈ R is the insulin ow injected for unit of weight (pmol/Kg/min).6.6 Regulator design 8. Linearization The mean Dalla Man et al. model. MPC model for prediction Considering the diculty of identifying an individualized model for each patient. the following 4 Linear Time-Invariant . which is nonlinear and time-variant. CHO ow assumed orally (mg/min). model with kempt (t. subcutaneous glucose d(t) ∈ R is the y(t) ∈ R is the measurable y = x13 = GM (t) (mg/dl). Linearizing the system around the basal steady-state point. Regulator design 201 8. we decided to use for prediction the mean Dalla Man et al. d(t)) ˙ y(t) = Cx(t) where (8. Hence. can be syn2 thetically described by x(t) = f (x(t). model. some steps are required in order to obtain a LTI 4 model as required by the LMPC formulation.

. B . d δd basal A ∈ R13×13 . u. with two inputs and one output. a constant value equal to the δu(t) = δu∗ (k). basal M= δf x. It consists of a zero-order hold (ZOH). consider the discrete- time system in Figure 8. for the whole time period of length last value of the input sequence: TS . MPC of Glycaemia in Type 1 Diabetic Patients 13th-order linear system. the The ZOH continuous-time system and the sampler connected in series. and M matrices are obtained as: A= δf x. MD and CD matrices.24) . M ∈ R13×1 Discretization After dening an appropriate sample time discretized TS . the system has been δx(k + 1) = AD δx(k) + BD δu(k) + MD d(k) In order to calculate (8.202 Chapter 8.23) δy(k) = CD δx(k) A D . d δx B= basal δf x. is obtained δ x(t) = Aδx(t) + Bδu(t) + M d(t) ˙ δy(t) = Cδx(t) where (8. u. B D .10. kTS ≤ t < (k + 1)TS (8. d δu B ∈ R13×1 .22) δx = x − xb δy = y − yb δu = u − ub where yb = Gb and A. u. holds.

TS MD = 0 eAη M dη. (8. a model order reduction and a balancing have been done.27) and equations (8. while the sampler gives as output kTS ≤ t < (k + 1)TS y ∗ (k) as: (8.M.25) y ∗ (k) = y(kTS ) Considering continuous-time Lagrange equation: (8. TS BD = 0 eAη Bdη.26) one has: AD = eATS .28) Sample time TS was xed equal to 15 minutes. Regulator design 203 δu*(k) TS δu(t) TS ZOH d*(k) d(t) A.6.CD δy(t) δy*(k) Figure 8. so obtaining a third order model described .8. CD = C (8.MD.B.10: ZOH method for system discretization d(t) = d∗ (k). balancing and Input-Output realization In order to reduce the complexity of the system.24). Order reduction.26) t t x(t) = eAt x(0) + 0 eA(t−τ ) Bu(τ )dτ + 0 eA(t−τ ) M d(τ )dτ .BD.C AD.25) and (8. t>0 (8.

the necessity of an observer is avoided. system (8. the two transfer functions have the same denominator. MPC of Glycaemia in Type 1 Diabetic Patients by the following transfer function ∆Y (z) = Ct (zI − At )−1 Bt ∆U (z) + Ct (zI − At )−1 Mt D(z) = = with N umM (z) N umB (z) ∆U (z) + D(z) Den(z) Den(z) (8.204 Chapter 8.29) can be given in the following state-space (nonminimal) representation (8.29) N umB (z) = b2 z 2 + b1 z + b0 N umM (z) = m2 z 2 + m1 z + m0 Den(z) = z 3 + a2 z 2 + a1 z + a0 Since. Finally. due to the balancing.30) δy(k + 1) = −a2 δy(k) − a1 δy(k − 1) − a0 δy(k − 2)+ + b2 δu(k) + b1 δu(k − 1) + b0 δu(k − 2)+ + m2 d(k) + m1 d(k − 1) + m0 d(k − 2) Note that.32) . by using an input-output representation of the system.31) ξ(k + 1) = AIO ξ(k) + BIO δu(k) + MIO d(k) δy(k) = CIO ξ(k) (8. the following input-output representation is obtained (8.

8. s are positive scalars.6. 0 0 0 0 0 0 0 1 0 (8. δu(k − 2) d(k − 1) d(k − 2) = 0 0 1 . Regulator design 205 where ξ(k) = δy(k) BIO (k) = b2 MIO (k) = m2 δy(k − 1) δy(k − 2) δu(k − 1) . r is the optimization horizon. u(·).35) Control law computation In order to derive the LMPC control law the following quadratic discretetime cost function is considered N −1 J(x(k). and y o (k) the desired output at time k. .33) −a2 −a1 −a0 b1 b0 m1 m0 1 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 AIO 0 0 0 0 0 0 (8. k) = i=0 q δy 0 (k + i) − δy(k + i) +s δy 0 (k + N ) − δy(k + N ) 2 2 + r u(k + i) 2 (8.36) where while N q.34) CIO = 1 0 0 0 0 0 0 (8.

. . N −3 C AN −2 B CIO BIO 0 IO CIO AIO BIO · · · IO IO N −1 N −2 CIO AIO BIO CIO AIO BIO · · · CIO AIO BIO CIO BIO CIO MIO 0 ··· 0 0 CIO AIO MIO CIO MIO ··· 0 0 . . . D(k) = d(k + N − 1) CIO AIO CIO A2 IO . . . . . .38) . N −3 C AN −2 M CIO MIO 0 IO CIO AIO MIO · · · IO IO CIO AN −1 MIO CIO AN −2 MIO · · · CIO AIO MIO CIO MIO IO IO (8. . Bc = . . . . d(k) .32) (8. .. . . . . . Ac = .206 Chapter 8. . . Mc = . . . . MPC of Glycaemia in Type 1 Diabetic Patients The evolution of the system can be re-written in a compact way as follows Y(k) = Ac x(k) + Bc U(k) + Mc D(k) where. δy(k + N ) . ..37) Y(k) = δy(k + 1) . C AN −1 IO IO CIO AN IO CIO BIO 0 ··· 0 0 CIO AIO BIO CIO BIO ··· 0 0 . U(k) = δu(k + N − 1) . . . . . considering (8. δu(k) . . .

Moreover. the cost function becomes J(x(k).8. following the Receding Horizon approach the control law is given by δu0 (k) = 1 0 ··· 0 U 0 (k) r has been xed at In order to tune just one parameter.36). in (8. U(k). matrices Q and R are dened as follows . Regulator design Ac . the solution of the optimization problem is (8. 1 and s = q. k) = = (δy 0 (k) − δy(k)) q(δy 0 (k) − δy(k))+ + (Y 0 (k) − Y(k)) Q(Y 0 (k) − Y(k))+ + U (k)RU(k) In the unconstrained case. Bc Mc 207 Matrices mula and are derived using the discrete time Lagrange for- i−1 ξ(k + i) = Ai ξ(k) IO + j=0 Ai−j−1 BIO δu(k + j) + MIO d(k + j) .41) = K0 Y (k) − Kx ξ(k) − KD D(k) where K = 0 K = x KD = −1 Bc QB c + R −1 Bc Q Bc QAc −1 (8. IO i>0 (8.40) −1 U 0 (k) = Bc QB c + R 0 Bc Q Y 0 (k) − Ac ξ(k) − Mc D(k) = (8.39) In this way.6.42) Bc QB c + R Bc QB c + R Bc QMc Finally.

41). BDIO (k) = δu(k − 2) d(k − 1) d(k − 2) 0 0 0 . MPC of Glycaemia in Type 1 Diabetic Patients Q = qI1 . In order to complete the regulator denition. times.208 Chapter 8. ξ(k) e D(k) that appear in (8. DDIO (k) = 0 1 0 0 0 0 0 0 0 (8. N = 240/TS = 16 sampling Generation of state vector ξ(k) In order to generate vector ξ(k) it is useful to dene the following dy- namic system. the scalar q and how to generate the vectors δY 0 (k).44) . Optimization horizon The optimization horizon has been xed for each virtual patient at 240 minutes (4 hours). where R = I2 I1 and I2 are identity matrices of appropriate dimensions.43) ξ(k) = CDIO ψ(k) + DDIO δym (k) δum (k) dm (k) ψ(k) = δy(k − 1) 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 δy(k − 2) δy(k − 3) δu(k − 1) . Since sample time is 15 min. it is necessary to dene the optimization horizon. whose aim is to store the previously measured values of inputs and outputs ψ(k + 1) = A DIO ψ(k) + BDIO where δym (k) δum (k) dm (k) (8.

0 0 0 CDIO 0 0 0 0 0 1 (8. . The warm-up procedure of the regulator consists in store the rst three measures of this variables in order to initialize the state vector ψ. that are used as inputs of the regulator.46) 1 = 112 − yb Generation of vector D(k) Vector D(k) represents the meals prediction along the optimization horiD(k) is equals to the amount zon. This is one of the key advantage of model predictive control with respect to other kind of feedback regulators such as the PID. that means the measured values of δy . (8. has been considered constant at 112 δY (k) = (y − yb ) 0 0 1 . Even if it is unrealistic to know in advance the real amount and time of . Regulator design = = 209 0 1 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 ADIO 0 0 0 . δu and d.6.8. that can be called meal announcement. . Generation of vector δY 0 (k) The glycemia reference vector mg/dl. Hence y0. is very useful since permits to react in advance with respect to distrubances on the controlled variable. . . This means that the i-th element of vector of carbohydrates (mg/min) that is supposed to be absorbed at time instant k + i − 1. 112 − yb . This information.45) Note that the inputs of such a system are measured variables.

all the states of the systems are measurables.47) diet(k + N − 1) 8. Nonlinear predictive control allows the use of nonlinear cost functions. that can be used in order to make the vector of predictable disturbances along the optimization horizon D(k) = diet(k) .6.2 NMPC Unconstrained LMPC with a quadratic cost function provides an explicit control law. We have decided to adopt a cost function related to the LBGI and HBGI indexes previously described in Section 8. we 1. consider the best unrealistic situation: In order to do this. A state feedback NMPC is proposed (we don't use an observer). . the cost function presents also .1. The model used in the synthesis of the NMPC is the full nonlinear continuous time model previously described in Section 8.5. the nonlinear model used in the synthesis of the NMPC describes exactly the system 2. MPC of Glycaemia in Type 1 Diabetic Patients future meals. we can suppose to know a nominal diet of the patient.210 Chapter 8. In order to minimize the control energy used along the optimization horizon.4. In the following a nonlinear model predictive control is proposed. The goal is to evaluate the (upper) performance limit of a predictive control algorithm in the blood glucose control problem. The simplicity of this solution is of great advantage in order to implement the controller on real patients. . (8.

the optimization horizon has been xed at N = 16 sampling times. a and b g = 0. The Finite Horizon Optimal Control Problem consists in minimizing. In order to tune just one of them. · · · . Regulator design 211 a term depending from the control action. . tk+i ). . with respect to control sequence u1 (tk ). τ ∈ [tk+i−1 .e. it is necessary to dene the scalars xed at 1. q and r. In order to complete the regulator denition. . uN (tk )) tk+N (8.44. a = 1.02 These values have been changed with respect to the original ones of the risk indexes in order to consider the fact that the cost function is a compromise between state and control requirements. As for the LMPC. .75. adopted in this case are g. i ∈ [1. u(τ ) = ui (tk ).8. . b = 9. the control inputs ui (tk ) are constrained to be piecewise constants. . u1 (tk ). The values of see [Magni & Scattolini 2004]. the following nonlinear cost function with horizon N J(xtk . q has been .6.48) = tk 10q(g((ln(Gp (τ )))a − b))2 + ru(τ )2 dτ Note that a full hybrid solution is developed. . N ]. uN (tk ) (tk = kTS with TS sam- pling time period). . i.

11): 1. Simultaneously patient receives an insulin bolus δ according to his personal in- . open-loop insulin therapy. switching time to closed loop and start time of regulation phase. The performance are tested on a clinical scenario. for example the simulation duration. the meal prole.7.0.11: Nominal scenario for adult patient 8.212 Chapter 8. At 18. It is a specic of all the parameters which characterize a given experiment.00 of Day 1. 8. Patient enters at 17.7 Clinical Scenario An in silico trial consisting of 100 patients is used to assess the performances of LMPC and NMPC controllers. The meal contains 85g of carbohydrates.00 of Day 1 patient has a meal of about 15 minutes of duration. 2. Simultaneously data gathering for warm-up phase starts. Microinfusor is programmed to inject basal insulin for the specic patient. MPC of Glycaemia in Type 1 Diabetic Patients Dinner 85g CHO + Bolo 450 400 Breakfast 50g CHO Warm-up Commutation Regulation Glycemia (mg/dl) 350 300 250 200 150 100 50 17 18 20 21:30 Day 1 23 24 2 4 6 Day 2 7:30 10 12 Figure 8.1 Nominal scenario Nominal scenario consists of the following phases (see gure 8.

Basing on the CVGA. Meal duration is considered about 2 minutes.00 of Day 1 regulation phase starts. the controllers are synthesized by xing all the parameters of their designs. restarts his normal insulinic therapy. in CVGA coordinates. At 7. However. it does not provide a numerical values that permits to quantify the quality of the performance.12 explains the control design procedure.8 Control design procedure LMPC and NMPC have respectively one parameter that must be tuned for each patient in order to achieve good performances in term of glycemia regulation. 6. At 23. Patient is discharged and 8. is a very useful tool in order to evaluate the outcome of the test. The test consists in the nominal scenario previously described.5. 5. 4. linear in the LMPC case. At this time the switching phase starts. a good performance index is the norm-innite of the distance.8. Experiment nishes at 12. Note . patient has breakfast containing 50g of CHO. these are applied to a subject realistic model of each patient. The CVGA. Figure 8.30 of Day 1. from the lower left corner of the grid. described in Section 8. that is δ = 85 · CR 3. nonlinear in the NMPC.30 of Day 2.00 of Day 2. Given a model.8. Then. once the control laws are determined. Control design procedure 213 sulin/carbohydrate ratio (CHO ratio o CR). In this phase the metrics for evaluation are not calculated. Control loop is closed at 21.

tunable parameters of the regulators (that means q or r) are changed.12: Control design procedure that the lines of level of this performance index are squared (see Figure 8.13). Whenever the results obtained are not satisfactory. for each patient (see [Tessera q 2007] for details).14 show an example of calibration curve for a patient in the LMPC case. In particular. An increase in the value of parameter causes a decrease in both maximum and minimum blood glucose (the reverse happens in NMPC case by changing parameter r). MPC of Glycaemia in Type 1 Diabetic Patients Figure 8.214 Chapter 8. Figure 8. a calibration procedure has been developed in order to obtain the best value of parameter q (or r). The arrow points the optimal value of the index at which an optimal q is associated. .

in terms of CVGA. the model used in the synthesis is exactly the nonlinear model used in the test. In particular.13: Performance index: norm-innite of the distance from the lower left corner of the CVGA 8. the CVGA of Figure 8.8.15 show the comparison between the NMPC with patient and the NMPC with an unique r calibrated for each r for all the population.9. NMPC is in the best situation: for each patient. Figure 8. The parameters q and r have been individually tuned. It is worth to note that clearly the rst solution gives better performances even if also the second one provide good results in terms of patients contained in the good zones (A and B). The gure show that. It is worth to note that LMPC uses a mean model of the population for the synthesis of the regulator of each patient.16 show the comparison between NMPC and LMPC. the unrealistic NMPC . Results 215 Figure 8.9 Results In this section the results obtained by applying LMPC and NMPC to the 100 virtual patient population are shown. Dierently.

Comparing with the LMPC one it is smoother in spite of a more or less equivalent glycaemia regulation.14: Example of calibration curve in LMPC case improves the performances.216 Chapter 8. The simulation ex- . 2007b]) have been proposed. In order to assess the perfor- mances of the proposed algorithm against interindividual variability. 8. continuous glucose monitoring is considered. insulin delivery and s.17 and regards the insulin prole.c. a linear model predictive control based and a nonlinear state feedback model predictive control (MPC) synthesized on a recently developed nonlinear in-silico model of glucose metabolism ([Magni et al. the feedback control of glucose concentration in Type-1 diabetic patients using s. but this improvement is not very signicant. In particular.10 Conclusions In this chapter. MPC of Glycaemia in Type 1 Diabetic Patients Figure 8. The real advantage of the NMPC is shown in Figure 8.c. an in silico trial with 100 virtual patients was carried out.

8.10. Conclusions 217 Figure 8.16: Comparison between LMPC with q individually tuned and NMPC with r individually tuned for all the patients .15: Comparison between NMPC with r individually tuned and r unique for all the patients Figure 8.

Juvenile Diabetes Research Foundation has launched a project to accelerate the development and adoption of a device for the closed loop control of blood glucose. An on-line optimization problem could also be a limitation for a real application. It is worth to note that in 2006.17: Comparison between LMPC and NMPC: glucose and insulin proles of a particular patient periments highlight the increased eectiveness of the meal announcement signal with respect to the linear MPC due to a more accurate nonlinear model. University . In particular.218 Chapter 8. Moreover. MPC of Glycaemia in Type 1 Diabetic Patients Figure 8. the following assumptions should be removed: perfect knowledge of the model parameters and state availability. one of the main advantages of a nonlinear approach is the possibility to use a nonlinear cost function based on the risk index dened in [Kovatchev et al. The obtained results encourage a deeper investigation along this direction. 2005].

Currently. experiments using the developed linear model predictive controller are in progress at the Charlottesville and Padova hospitals.8. . Conclusions 219 of Pavia collaborates in this project joined with the University of Virginia and the University of Padova.10.

.

given two sets is the point-to-set ζ ∈ IRn A A ∈ IRn .t2 ] be a signal sequence dened from time t1 to time t2 . IR≥0 . w. dist(A. Let IR. where w supk≥0 {|wk |} w[τ1 . whose values belong to a compact set W ⊆ IRm sup by MW . The symbol id represents the identity function from is the composition of two functions Given a set distance from IR to IR. .τ2 ] w supτ1 ≤k≤τ2 {|wk |} wk denotes the values that the sequence takes in correspondence to the index k. In order to simplify the notation.Chapter 9 Appendix of the Thesis This section provides the main notations and denitions used in the thesis. B) inf {|ζ|A . B ∈ IRn . let denote with σ (A) ¯ the maximum singular A. ζ ∈ B}. |ζ|A to inf {|η − ζ| . W inf and inf w∈W {|w|}. let Given a signal w[t1 . while W supw∈W {|w|}. η ∈ A} while. while γ1 ◦γ2 to γ1 and γ2 from IR IR. sequences w. A ⊆ IRn . the integer and the non-negative integer sets of numbers. the subscript of the sequence is omitted. respectively. The Euclidean norm is denoted as norm. Given a matrix value of |·| while | · |∞ denotes the innity- A ∈ IRn×n . when it is inferrable The set of is denoted Moreover from the context. Z and Z≥0 denote the real. the non-negative real.

dk . β(s. Appendix of the Thesis A ⊆ IRn and B ⊆ IRn with B ⊆ A. for each xed t ≥ 0. s : Z≥0 → IR≥0 . t) as is of class s ≥ 0. n x ≥ y. n then the Pontryagin dierence set C=A B {x∈IRn : x+ξ∈A. positive denite and strictly increasing. x ≥ y ⇐⇒ xi ≥ yi . for each xed is of Denition 9. C is denotes the boundary of Given two sets dened as int(A) denotes the interior of A∈IRn .4 (Upper limit) the upper limit is dened as Given a bounded function limt→∞ s(t) t≥0 τ ≥t inf sup s(τ ). x ∈ B}. while the Minkowski sum .2 (K∞ -function) if it is a A function as γ : IR≥0 →IR≥0 is of class K∞ K-function and γ(s) → +∞ s → +∞. η and of radius ρ . k ≥ 0 (9. is denoted as B(η. Denition 9. i = 1.3 (KL-function) class A function KL if. Given a closed set The dierence between two given sets is denoted by A\B {x : x ∈ A. · · · . wk ). and x. x=ξ +η} η ∈ IRn and a positive scalar Given a vector centered in ρ ∈ IR>0 . ρ) {ξ∈IRn : |ξ −η|≤ρ}. y ∈ IRn . t) → 0 t → ∞. B ∈IRn .1) . ∂A A. The shorthand For B(ρ) is used when the ball is centered in the origin.1 (K-function) γ : IR≥0 →IR≥0 is of class K (or a K-function) if it is continuous. ·) is decreasing and β(s. ∀ξ∈B}. the closed ball set is dened as S=A⊕B {x∈IR : ∃ξ ∈A. β(·. / A while A ⊆ IRn . Consider a discrete-time autonomous nonlinear model xk+1 = F (xk . Denition 9. η∈B.222 Chapter 9. β : IR≥0 × Z≥0 → IR≥0 K. A function x y means the negation of Denition 9.

0.8 (ISS in IR n and Ξ with respect to A) Ξ Given a compact set Ξ⊂ including the origin as an interior point. d. x.6 (0-AS in Ξ) to be Ξ ⊂ IRn and including the origin as an interior point. ¯ Denition 9. x. all d ∈ D(x). 0)| ≤ β(|¯|. w) ∈ Ξ x ∈ Ξ. k).1) and if there exists a K∞ - γAG such that for all initial values x ∈ Ξ. w). models a class of uncertainty which depends on the state (for each while the set D(x) is closed). all d ∈ D ¯ and all w ∈ W.1) with initial state is denoted by x0 = x ¯ and disturbance sequences d and w x(k. ∀k ≥ 0. is said to be ISS in with respect to set A (compact-ISS). ¯ x x (9.2) Denition 9. wk ∈ W ⊆ IRq models a class of disturbance. if Ξ is a RPI set for (9.1) with d ∈ MD w ∈ MW . d. Given a compact set Denition 9. d.7 (AG in Ξ) (AG) property in function one has : Given a compact set Ξ ⊂ IRn including the origin as an interior point.1) satises the Asymptotic Gain Ξ. The transient of the system (9. w ∈ W. x.1) with d=0 w=0 is said 0-AS in Ξ. ∀¯ ∈ Ξ. if and a Ξ is a RPI set for (9. d.1).1) and if there exist a KL-function β γ2 such that K-function |x(k. ¯ A set Denition 9.223 xk ⊆ R n dk ∈ D(x) ⊆ IRp x where is the system state.5 (RPI set) and all Ξ ⊆ IRn is a Robust Positively Invariant for all (RPI) set for the system (9. if Ξ is positively invariant and if there exists a KL-function β such that |x(k. w)| ≤ γAG (||w||). limk→∞ |x(k. the system (9. w)|A ≤ β(|¯|A . the system (9. the system (9. k) + γ2 (||w||) ¯ x . x. if F (x.

1) with d ∈ MD and w ∈ MW satises the LpS (Local practical Stability) property if there exists a constant c≥0 ε > 0. Denition 9.10 (LS with respect to A) and d1 ∈ MD w ∈ MW satises the LS (Local Stability) property with respect to set A.1) is an equilibrium point and c = 0. x. the system is said to satisfy the UAG (Uniform Asymptotic Gain) property. Chapter 9. Ξ ⊂ IRn and including Denition 9. x. w)| ≤ γ2 (||w||) + c + ε ¯ for all x∈Ξ ¯ with |¯| ≤ ν . ν ) such that |x(k.9 (LpS) such that for each System (9. Appendix of the Thesis for all and Denition 9. w)| ≤ c + ε ¯ for all k ≥ 0. if for each ε > 0.1) with Denition 9. d.12 (UAG in Ξ with respect to A) Given a compact set . there exists a δ>0 such that |x(k. d. there exists a δ>0 such that |x(k. d. If the origin of system (9. if satises the UpAG (Uniform practical Asymptotic Gain) property in is a RPI set for (9.224 x∈Ξ ¯ k ≥ 0. x and all k ≥ T.1) with d ∈ MD1 c≥0 w ∈ MW Ξ. x. system (9. System (9. all |¯| ≤ δ x and all |wk | ≤ δ . ∃T = T (ε. w)|A ≤ ε ¯ for all k≥0 for all |¯|A ≤ δ x and all |wk | ≤ δ .1) and if there exist a constant and a Ξ K-function γ2 such that for each ε>0 and ν > 0.11 (UpAG in Ξ) Given a compact set the origin as an interior point.

wk ). system (9.225 Ξ ⊂ IRn and including the origin as an interior point.1) with d ∈ MD w ∈ MW Ξ ε>0 satises the UAG (Uniform Asymptotic Gain) property in Ξ. Ξ ⊆ X κ(xk ) ∈ U . d. ∃T = T (ε. wk ∈ W ⊆ IR models a class of disturbance.3) with implies uk = κ(xk ). ¯ Given u = κ(x). dk .13 (RPIA set) a control law Consider system (9. ν ) such that |x(k.3) with x0 = x. wk ∈ W k ≥ 0.1) and if there exist a and K-function γ2 such that for each ν > 0. all if x ∈ Ξ ¯ and all xk ∈ Ξ and for all dk ∈ D1 (x. . x and all k ≥ T. u) ⊆ IRp q models a class of uncertainty which depends on the state and the control input (for each while x and u the set D1 (x. is a Robust Positively Invariant Admissible (RPIA) set for the closed-loop system (9. uk . dk ∈ D1 (x. if is a RPI set for (9.3) xk ∈ X ⊆ R n is the system state. Denition 9. x. w)|A ≤ γ2 (||w||) + ε ¯ for all x∈Ξ ¯ with |¯|A ≤ ν . k ≥ 0 where (9. uk ∈ U ⊆ IRm is the current control vector. Consider now a discrete-time nonlinear model xk+1 = f (xk . u) is closed). u).

.

Qin. A characteriza- tion of integral input-to-state stability. R. Alamo. pages 389404. J. Vella. . 2003. Frank. Springer. D. Sontag and Y. Nonlinear Predictive Control and Moving Horizon Estimation . 2005] T. 2000] D.R. 43. Di Camillo. An articial endocrine pancreas. B. Z. G. Ewart. 2003] R. Limon and E. Allgöwer. vol. Wright.Bibliography [Alamo et al. 5. pages 710714. B. 284. pages E55 E69. Albisser. 45. Advances in Control. vol. pages 10821097. Camacho. 2000. [Albisser et al. [Alamo et al. Automatica. Basu. no. D. Constrained min-max predictive control: modications IEEE of the objective function leading to polynomial complexity. Botz and W.K. Muñoz de la Peña. pages 391449.Endocrinology and Metabolism. [Basu et al. T. P. Minmax MPC using a tractable QP problem. Toolo. T. 50. C. 4. A. Cobelli. Wang. D. 1974. no.G. Davidovac. In Paul M. 1974] A. 1999] F.S. 2007] T.F. D. Muñoz de la Peña and E. E. American Journal of Physiology . Zingg. vol. 2007.An Introductory Overview. Diabetes. Ramirez. F. Badgwell. Use of a novel triple-tracer ap- proach to assess postprandial glucose metabolism. A. IEEE Transactions on Automatic Control. [Angeli et al. B. 1999.J.M. 23. J. A. D. editeur. Alamo. Rawlings and S. 2005. S. [Allgöwer et al. Shah. Leibel. Camacho. Transactions on Automatic Control. Basu. no. pages 693700. Angeli. Rizza and C. vol.

F. [Bemporad 1998] Alberto Bemporad. no. Recent Progress in Hormone Research. Automatica. F. Springer. [Blondel & Tsitsiklis 2000] V. Reducing conservativeness in predic- tive control of constrained systems with disturbances. vol.N. Automatica. F. vol. [Bergman & Urquhart 1971] R. Bravo. 1971. Alamo and E. Camacho and C. Distributed Model Predictive Control. 27. 36. Bordons. vol. . Camacho. Model Pre- [Camponogara et al. Krough and S. H. Automatica. IEEE Control Systems Magazine. 2000. 2005. 2002] E. Urquhart. Blondel and J. Limon. On the minmax reachability of target set and target tubes. 9. T. D. pages 4452. Min-max control of constrained uncertain discrete-time linear systems. pages 15831589. dictive Control. Bemporad. 2004. pages 233247. Tsitsiklis. Camponogara. [Camacho & Bordons 2004] E. 48. A survey of computational complexity results in systems and control. [Bravo et al. The pilot gland approach to the study of insulin secretory dynamics. pages 12491274. 2003] A. pages 16001606. B. Talukdar. 7. [Bertsekas & Rhodes 1971] D. D. vol. 2005] J. Rhodes.D. Jia. Borrelli and M. pages 583605. 41. 2002. Bergman and J. 1998. 1971. N. vol. 2003. Morari. Bertsekas and I. IEEE Transactions on Automatic Control. M.228 Bibliography [Bemporad et al. On the computation of invariant sets for constrained nonlinear systems: An interval arithmetic approach.

Serrani. 1996] L. Automatica. 1997. A. Chandler. Bahamas. In Proc of. 2001] L. 1998. pages 12051217. [Chisci et al. Dual receding horizon control of constrained discrete-time systems. Chen. 1998. Chen and L. vol. no. 37. C. 1998] H. Chen. on Decision and Control. 2000] P. of the IEEE Conf. [Chen & Shaw 1982] C. Pachter. pages 10191028. Zappa. Scherer and F. Chen and A. Rasmussen and M. [Chen et al. [Chen & Allgöwer 1998] H. 3. . A game theo- retical approach to nonlinear robust receding horizon control of constrained systems. Greece. pages 349352. vol. [Chisci et al. [Chen & Serrani 2004] X. Remarks on ISS and For- mation Control. Scherer and F. 2. On Receding Horizon Feed- back Control. Mosca. Chisci. In Proc. Rossiter and G. UAV cooperative path planning. Lombardi and E. Chisci. Corfu. J. Automatica. A quasi-innite horizon nonlinear model predictive control scheme with guaranteed stability. 34. 2000. 4. Allgöwer. Allgöwer. Chen and F. CO. vol. C. In Proc. DYCOPS 5. 18. Denver. vol. Shaw. Allgöwer. pages 177182. In Proc. of the AIAA Guidance. 1997] H. 1982. and Control Conference and Exhibit. S. Navigation. A. no.Bibliography 229 [Chandler et al. pages 278285. [Chen et al. 2001. the IEEE American Control Conference. A robust model predictive control scheme for constrained linear systems. Systems with Persistent Disturbances: Predictive Control with Restricted Con- straints. European Journal of Control. Automatica. W. 1996. C.R. 2004. W.

Camilleri and C. Raimondo. 2003. Lall. vol. of the Independent Stability of Interconnected Systems. 2006] C. vol. no. IEEE American Control Conference. M. vol. Rüer and F. A. 19. 2007a] C. Distributed Control Design for Spatially Interconnected Systems. . pages 24722478. Mathematics of Control. Dullerud. Rizza and C.R. pages 3717 3722. of the 44th IEEE Conf.230 Bibliography [Cogill & Lall 2004] R. An ISS small gain theorem for general networks. Dalla Man. no. Control Design for TopologyIn Proc. D'Andrea and G. vol. on Decision and Control. 2006. Meal Simulation Model of the Glucose-Insulin System. pages 93122. B. Boston. 10. Signals. D. 2007. Dalla Man. vol. 2007b] C. IEEE Transactions on Automatic Control. Simulation Software of Meal Glucose Insulin Model. 2007. R. Dalla Man. IEEE Transactions on Biomedical Engineering. 3. R. A Sys- tem Model of Oral Glucose Absorption: Validation on Gold Standard Data. Dashkovskiy. Cobelli. In Proc. 12. Cobelli. Dashkovskiy. no. IEEE Transactions on Biomedical Engineering. [Dalla Man et al. M. [Dashkovskiy et al. 2007] S. A small-gain type stability criterion for large scale networks of ISS systems.A. Journal of Diabetes Science and Technology. Wirth. 2005. pages 323 330. S. [D'Andrea & Dullerud 2003] R. 9. no. Wirth. 48. Cogill and S. B. 1. 53. 54. R. and Systems (MCSS). no. Cobelli. GIM. [Dalla Man et al. Rüer and F. 2005] S. [Dashkovskiy et al.S. 2. 2007. 2004. pages 14781495. [Dalla Man et al. Rizza and C. pages 17401749. pages 56335638. E.

Progress in Systems and Control Theory. [De Nicolao et al. L. 1998. Scattolini. In F. 2000. Switzerland. De Nicolao. 1998. [De Nicolao et al. In NMPC Workshop . IEEE Transactions on Automatic Control. Scattolini. L. 1996. vol. 43.Bibliography 231 [De Nicolao et al. 41. Sta- bilizing receding-horizon control of nonlinear time-varying systems. vol. De Nicolao. L. Stability and robustness of nonlinear receding-horizon control. pages 10301036. 1998b] G. de Oliveira Kothare and Contractive Model Predictive Control for Constrained Nonlinear Systems. pages 10301036. [De Nicolao et al.Assessment and Future Directions. L. 1996] G. Scattolini. AC-43. M. no. Scattolini. Zheng. L. Magni and R. 1998. Nonlinear Model Predictive Control. Automatic Control. [de Oliveira Kothare & Morari 2000] S. Ascona June 3-5. Magni and R. 1998c] G. Magni and R. pages 322. [De Nicolao et al. 7. Sta- bilizing Receding-Horizon Control of Nonlinear Time-Varying Systems. Morari. IEEE Transactions on Automatic Control. IEEE Transactions on. vol. pages 7790. De Nicolao. On the Robustness of Receding-Horizon Control with Terminal constraints. IEEE Transactions on Automatic Control. Stabil- ity and robustness of nonlinear receding-horizon control. pages 451453. pages 10531071. . Magni and R. 2000. Birkhauser Verlag. Scattolini. De Nicolao. L. 2000] G. 1998a] G. Allgöwer and A. Magni and R. De Nicolao. Ascona. editeurs.

DeHaan and M. IEEE Transactions on Automatic Control. C. [Dunbar & Murray 2006] W. pages [Fontes & Magni 2003] F. L. on Decision and Control. pages 14781491. pages 19952002. pages 42744279. Dua. A. M. 10. 190206. pages 17501755. no. Min-Max Model Predictive Control of Nonlinear Systems Using Discontinuous Feedbacks. 8. 9. Doyle III and E. Adaptive Robust MPC: A minimally-conservative approach. 2004. Polycarpou. Foss. no. 2-3. vol. C. 53. Model- Based Blood Glucose Control for Type 1 Diabetes via Parametric Programming. F. Imsland. Findeisen.232 Bibliography [DeHaan & Guay 2007] D. of the IEEE Conf. . vol.J. Coopera- tive Control of DiscreteTime Agents with Delayed Information Exchange: a RecedingHorizon Approach. 2006. A. B. 2003. 2006] P. [Dua et al. Franco. T. Guay.N. 2007. Distributed receding horizon control with application to multi-vehicle formation stabilization. of the IEEE Conf. Murray. Allgower and B. pages 39373942. Dunbar and R. pages 549558. 2006. Bahamas. Automatica. Bahamas. IEEE Transactions on Biomedical Engineering. 2004. 2003] R. M. 2004] E. on Decision and Control. vol. In Proc. [Findeisen et al. 2003. In Proc. Receding Hori- zon Control of Multi-Vehicle Formation: a Distributed Implementation. 48. [Franco et al. 42. B. [Dunbar & Murray 2004] W. vol. European Journal of Control. Fontes and L. F. In American Control Conference 2007. Murray. State and output feedback nonlinear model predictive control: An overview. Magni. Pistikopoulos. Dunbar and R. Parisini and M. no.

Bequette. Int. on Automatic Control. pages 324338. Molnar. 2008] E. 23. Gopinath. pages 425435. Magni. pages 703 710. 1987. R. Computer and Biomedical Research. 1987. Gatewood.Bibliography 233 [Franco et al. Parisini and M. pages 1527.C. vol. Simulation Studies of blood-glucose regulation: Eect on intestinal absorption. vol. Issues in the design of a multirate model-based con- . In IASTED International Conference on Control and Applications. IEEE Transactions on Automatic Control. On the ro- bustness of optimal regulators for nonlinear discrete-time systems.W. Yu. Cooperative Constrained Control of Distributed Agents With Nonlinear Dynamics and Delayed Information Exchange: A Stabilizing Receding-Horizon Approach. Raimondo. Robustness of nonlinear state feedback.a survey. Lin. pages 8186. vol.J. Parisini. C. Kaufman and C. 2000. Rosevear and G. H. T. T. M. 1995] R. [Glad 1987] S. Ackerman. B. M. Polycarpou and D. 2008.D. Gao and Y. 1968] L. [Gatewood et al. Franco. Control. Glad.W. [Gopinath et al. 2006. [Geromel & Da Cruz 1987] J. Design and Stability Analysis of Cooperative RecedingHorizon Control of Linear DiscreteTime Agents. Journal on Robust and Nonlinear [Franco et al. L. Polycarpou. no. Franco. On Equivalent Notions of Input-to- State Stability for Nonlinear Discrete Time Systems. [Gao & Lin 2000] K. AC-32. J. Automatica. IEEE Trans. 1. 1968. 53. J. E. 2006] E. T. Da Cruz. vol. Geromel and J. Roy. 2.

10. 149. Messina. pages 540546. 2002. pages 523533. 40. Tuna and A. R. pages 17291738. R. J. M. IEE Proceedings . E. 50. 2006] P. 2004. E. [Gyurkovics & Takacs 2003] Eva Gyurkovics and Tibor Takacs. 50. Maciejowski. E. J. All is Not Lost.M. M. 2005. Optimization over state feedback policies for robust control with constraints. S. vol. Grimm. Grimm. J. no.J. pages 546558. pages 277289. IEEE Transactions on Automatic Control. pages 18561870. Grimm. 10. Teel. vol. Kerrigan and J. vol. E. Nominally Robust Model Predictive Control With State Constraints. Model Predictive Control: For Want of a Local Control Lyapunov Function. Tuna and A. [Gyurkovics 2002] Eva Gyurkovics. 4. J. Grimm. [Grimm et al. In 42nd IEEE CDC. Biotechnology Progress. Automatica. pages 14131418. 2003] G. no. vol. Goulart. Teel.C. M. IEEE Transactions on Automatic Control. S. M. [Grimm et al. vol. [Grimm et al. . 2003. 2004] G. Messina. no. Teel. 2007] G.Control Theory and Applications. E. R. 2003. Tuna and A. Teel. [Goulart et al. 2006.234 Bibliography troller for a nonlinear drug infusion system. Quadratic Stabilisation with H∞ -Norm Bound of Non-Linear Discrete-Time Uncertain Systems with Bounded Control. Systems & Control Letters. Receding Horizon H∞ Control for Non- linear Discrete-Time Systems. Examples when nonlinear model predictive control is nonrobust. S. 52. 42. 2007. vol. Automatica. R. Nominally Robust Model Predictive Control with State Constraints. Tuna and A. pages 318332. 1995. Messina. 11. vol. S. 2005] G. Messina. [Grimm et al.

M. Jimenez. Canonico. pages 905920.cfm?page_id=101982. V. Massi-Benedetti.Bibliography 235 [Homko et al. L. Hovorka. Physiological Measurement. Nonlinear model predic- tive control of glucose concentration in subjects with type 1 diabetes. [JDRF Website ] JDRF Website. Nesic and P. 4. vol. Input-to-State Stability for Discrete-Time Nonlinear Systems. M. [Ioannou 1986] P. vol.J. 2001. IEEE Transactions on Automatic Control. Pharma- cokinetic and metabolic eects. pages 291 298. M. D. 2004.R.E. pages 776783. M. Jiang and Y. 2005] S. 2001. Schaller. H. R. Wilinska. Ioannou.jdrf. Pieber. C. IEEE Transactions on Automatic Control. 50. vol. 2005. [Jiang & Wang 2001] Z. on Automatic Control. Jadbabaie and J. vol. pages 16811697. Chassin. IEEE Trans. no. Schaupp. [Hovorka et al. 1986. J. Diabetes Care. Decentralized Adaptive Control of Intercon- nected Systems. [Jadbabaie & Hauser 2001] A. http://www.-P. pages 20272031. 31. vol. Deluzio. T. 7. Unconstrained Receding-Horizon Control of Nonlinear Systems. L. 26. 2004] R. James. Huang. Boden. Comparison of Insulin Aspart and Lispro. pages 857 869.W. Vering and M. Haueter. Kolaczynski and G. Hauser. vol. Dower. . 2003. 5. 25. Orsini Federici. [Huang et al. A unied Approach to Controller Design for Achieving ISS and Related Properties. 2003] C Homko. T. A.C.org/index. 37. U. Wang. Automatica. no.

H. A.-P. Praly. [Jiang et al. Teel and L. pages 21292136. of the 41st IEEE Conf. pages 265293. pages 395405. Kerrigan and J. Kerrigan. 2004. pages 5321 5326. Optimal control of constrained.. In Proc. 52. 14. Wildanger. Teel. Automatica. P. B. Lin and Y. R. Systems & Control Letters. [Kerrigan & Maciejowski 2004] E. Gilbert. 2002. on Decision and Control. International Journal of Robust and Nonlinear Control. Scholze. 2004. no. Foley. D. Y. pages 395413.M. 2004. and Systems. Jin. Optimal. pages 95120. [Keerthi & Gilbert 1988] S.R. Feed- back min-max model predictive control using a single linear program: robust stability and the explicit solution. A geometric approach to reachability computations for constrained . Th. Lygeros and J. vol. 4. J. [Jin et al. Small-gain theorem for ISS systems and applications. C. Boston. C. Maciejowski. vol. 2004] Z. Optimiz. 1994] Z. Murray. 7. [Kerrigan & Mayne 2002] E.236 Bibliography [Jiang et al. S. Wang. 2004. The Second Gener- ation Caltech Multi-vehicle Wireless Testbed.C. Mathematics of Control. 57. Q. Jiang. 40. Lammers. Held and R. Jiang. vol. Keerthi and E. S. G. [Kellett & Teel 2004] C. 2002] E. vol. pages 1552 1557. Kellett and A. 2004] Z. Maciejowski. M. E. Mayne.-P. M. M. J. 1994. Nonlinear small-gain theorems for discrete-time feedback systems and applications. Waydo. [Kerrigan et al. innite- horizon feedback laws for a general class of constrained discrete-time systems. Appl. 1988. vol. of the IEEE American Control Conference. Kerrigan and D. piecewise ane systems with bounded disturbances. Smooth Lyapunov func- tions and robustness of stability for dierence inclusions. Signals. M. In Proc.

no.W. pages 323328. Kevicky. vol. [Kevicky et al. of the IEEE American Control Conference. In Proc. Kovatchev. F. Bahamas. Balakrishnan and M. Prentice Hall. 7. pages 15921597. 2005. vol. Yoon. C. 32. Jadbabaie and C. The articial pancreas: How sweet engineerSystems & Control ing will solve bitter problems. [Klono 2007] D. De Persis. Balas. 2004. pages 49214926. J. 2004b] T. F. 2002. Hierarchical Design of Decentralized Receding Horizon Controllers for Decoupled Systems. M. W. pages 293303. pages 7281. Nonlinear systems. A. Klono. Brayman and A. 1996. 2007. Ro- bust constrained model predictive control using linear matrix inequalities. pages 849862.Bibliography 237 discrete time systems. V. T. In Proc. Letters. J. [Khalil 2001] H. Borrelli and G. Kim. Quantifying Temporal Glucose Variability in Diabetes via Continuous Glucose Monitoring: Mathematical Methods and Clinical Application. Borrelli and G. Kevicky. 2006] J. . of Diabetes Science and Technology. [Kim et al. 2001. pages 13611379. 6. A Study on Decentralized Receding Horizon Control for Decoupled Systems. Boston. J.S. vol. K. McCall. Khalil. 1.P. Diabetes Technology & Therapeutics. no. 2004. Clarke. 55. Morari. [Kevicky et al. In Proc. [Kothare et al. Balas. 2004a] T. 1996] M. Breton. 4. [Kovatchev et al. 15th World Congress IFAC. 2005] B. Input-to-state stable nite horizon MPC for neutrally stable linear discrete-time systems with input constraints.L. on Decision and Control. V. Kothare. vol.K. 2006. of the IEEE Conf.

on Decision and Control. Muñoz de la Peña. 2008] M. Heemels.238 Bibliography [Kovatchev et al. M. 2008. 2002. Lazar. M. 33. 2005] M. S. Las Vegas.M. Workshop on Assessment and Future Directions of NMPC. Cassandras. Raimondo. G. Lazar. Kovatchev. [Lazar et al. Lazar. In Proc. Yu.H. of the IEEE Conf. vol. 1997. 2006. H. 2005. Freudenstadt-Lauterbad. M. A Receding Horizon Approach for Solving Some Cooperative Control Problems. pages 327334. In Proc. on Decision and Control. The Netherlands. Lee and Z. D. Cassandras. [Lazar 2006] M. [Lee & Yu 1997] J. G. Automatica. pages 37603765. On the Stability and Robustness of Non-smooth Nonlinear Model Predictive Control. Weiland. of the IEEE Conf. Heemels and T. H. Hawaii. Germany. 2008. Bemporad and S. Worst-case formulations of model predictive control for systems with bounded parameters. [Lazar et al. In Silico Testing and in Vivo Experiments with Closed-Loop Control of Blood Glucose in Diabetes. [Li & Cassandras 2002] W. D. [Li & Cassandras 2003] W. W. Patek and C. PhD thesis. Systems & Control Letters. vol. Model predictive control of hybrid systems: Stability and robustness. 57. pages 763781. Li and C. P. 2008] B. W. Stability properties of a Cooperative Receding Horizon Controller. Breton. . Li and C. 2003. On Input-to-State Stability of Min-max Nonlinear Model Predictive Control. Alamo. In Int. pages 492497. Eindhoven University of Technology.P. A. Cobelli. pages 3948.

Automatica. Limon. 2002. Salas and EF Camacho. Alamo. [Limon et al. vol. 2008] D. Input-to- State Stable MPC for Constrained Discrete-Time Nonlinear Systems with Bounded Additive Uncertainties. Robust MPC of constrained nonlinear systems based on inter- val arithmetic. 51. Muñoz de la Peña. 2005] D. on Decision and Control. Camacho. IEE Proceedings-Control Theory and Applications. Raimondo. pages 364369. Input-to-State Stability: an . 5. on Decision and Control. In Proc. Camacho. F. T. of the IEEE Conf. 2006b] D. 2002. In Proc. page 325. J. [Limon et al. T. Alamo and E. Salas and E. Limon. D. G. IEEE Transactions on. Camacho. F. Limon. Limon. Alamo and E. Automatic Control. 2002b] D. Limon. pages 46194624. Camacho. vol. 2005. pages 29052910. F. Limon. F. Alamo and E. Alamo. pages 797 803. Stability Analysis of Systems with Bounded Additive Uncertainties Based on Invariant Sets: Stability and Feasibility of MPC. Bahamas. F. Cassandras. On the stability of constrained MPC without terminal constraint. Camacho. T.Bibliography 239 [Li & Cassandras 2004] W. 2004. [Limon et al. F. [Limon et al. 42. T. 152. D. T. pages 832836. 2006. M. vol. no. M. T. of the 41st IEEE Conf. Input to State stability of min-max MPC controllers for nonlinear systems with bounded uncertainties. Stability Properties of a Receding Horizon Controller for Cooperating UAV's. 2006a] D.M. Li and C. [Limon et al. F. J. 2006. Alamo. 2002a] D. In ACC02. [Limon et al. Bravo and E. Bravo.

September 5-9. IEEE Transactions on Automatic Control. 2000. 36. Sontag and Y. vol. Model Predictive Control of Continuous-Time Nonlinear Systems with Piecewise Constant Control. Pavia. [Löfberg 2003] J. [Maciejowski 2002] J. Department of Electrical Engineering. Liu. straints. Magni and R. 1. 2003. In International Workshop on Assessment and Future Directions of Nonlinear Model Predictive Control. Sweden. [Linkens & Mahfouf 1994] D. K. IEEE Transactions on Automatic Control. pages 124160. vol. 2002. Generalized pre- dictive control (GPC) in clinical anaesthesia.240 Bibliography Unifying Framework for Robust Model Predictive Control. pages 900906. Advances in ModelBased Predictive Control. 1996. PhD thesis. pages 16691674. Polycarpou. pages 7695. Linköping University. Predictive Control: With Con- [Magni & Scattolini 2004] L. no. Löfberg.M. 1. [Lin et al. 2003] Y. E. Arkun. 1994. Control and Optimization. [Liu et al. Wang. Italy. Automatica. pages 527540. Mahfouf. vol. A. Stability Analysis of M-Dimensional Asynchronous Swarms with Fixed Communication Topology. Lin. Quasi-Min-Max MPC algorithms for LPV systems. Lu and Y. A smooth converse Lyapunov theorem for robust stability. [Lu & Arkun 2000] Y. Maciejowski. vol. Passino and M. 2004. 34. SIAM J. Scattolini. D. Linkens and M. vol. 2003. Minimax approaches to robust model predic- tive control. Prentice Hall. 48. 2008. . 49. 1996] J Y.

Magni. Raimondo and R. Allgöwer. vol. pages 13511362. Robust Model Predictive Control of Nonlinear Discrete-Time Systems. R. 2007. [Magni & Scattolini 2006] L. Robustness and Ro- bust Design of MPC for Nonlinear Discrete-Time Systems. Automatica. pages 429435. Scattolini and F. Scattolini. pages 239254. pages 229246. Magni and R. H∞ control problem. volume LNCIS 358. Findeisen et al. 32. vol. [Magni & Scattolini 2007] L. pages 12311236. Magni. vol. Scattolini. 13. G. van der Schaft. Sepulchre. IEE Proceedings . 2006. 2006a] L. 42. In R. Control Design for Nonlinear Systems: Trading Robustness and Performance with the Model Predictive Control Approach. Magnani and R. pages 241245. M. pages 333339. H.): Assessment and Future Directions. [Magni et al. A receding-horizon approach to the nonlinear tomatica. 2001. Nijmeijer and A. [Magni et al. (Eds. Re- gional Input-to-State Stability for Nonlinear Model Predictive Con- . A stabilizing model-based predictive control for nonlinear sys- tems. [Magni et al. 2001b] L. 2003] L. D. 37. Springer-Verlag Berlin Heidelberg. 2005. Systems & Control Letters. Magni. Magni and R.Control Theory & Application. De Nicolao. 2003. Magni. L. G. [Magni & Sepulchre 1997] L. vol. 1997. Stability margins of nonlinear receding horizon control via inverse optimality. 37. International Journal of Robust and Nonlinear Control. vol. Scattolini. Magni and R. Magni and R. Scattolini. Au- [Magni et al. Automatica. J.Bibliography 241 [Magni & Scattolini 2005] L. 2001a] L. 2001. Stabilizing decen- tralized model predictive control of nonlinear systems. De Nicolao. Scattolini.

1. Breton. Kovatchev and C. D. 2006] G. Cobelli. 2008a] L.E. [Magni et al. July 6-11. 60/984. Magni. [Magni et al. 2006. An Improved PID Switching Control Strategy for . C. Kovatchev. De Nicolao. pages 630635. Seoul. pages 4246 4251. Barolo. 2006. vol. In 17th IFAC World Congress. 2007b] L. C. Zissery and D. B. Magni. Magni. Seborg. [Magni et al. Serial No. no. M. Input-to- State Stability for Nonlinear Model Predictive Control. [Magni et al. 2008b] L. 2008. 9. S. DM Raimondo and R. pages 48364841.M. 2007a] L. Dalla Man. [Magni et al. Bossi. vol. G. U. Model Predictive Con- trol of Type 1 Diabetes: An in Silico Trial. Raimondo. no. Raimondo. 2008. M. Raimondo. De Nicolao and D. C. L. Cobelli. Cobelli and B. 4.S. Scattolini. pages 804812. Magni. 2006b] L. D. M. C. 2006 45th IEEE Conference on. G. Dalla Man. B. 2007. Journal of Diabetes Science and Technology. H. Model Predictive Control Based Method for Closed-Loop Control of Insulin Delivery in Diabetes Using Continuous Glucose Sensing.956. Model Predictive Control of glucose concentration in subjects with type 1 diabetes: an in silico trial. 2007. Dalla Man. vol. pages 15481553. 51. DM Raimondo. Provisional Application Patent. [Marchetti et al. 2. G. Marchetti. In Decision and Control. no. G. 6. Kovatchev and C. Magni. Jovanovicy. M. IEEE Transactions on Automatic Control. De Nicolao. Evaluating the ecacy of closed-loop glucose regulation via control-variability grid analysis (CVGA). Patek. L. Korea. de Nicolao. Journal of Diabetes Science and Technology.242 Bibliography trol.

23. 41. Nesi¢ and D. vol. pages 16231633. vol. 2000] D. 2005] P. [Mayne et al. 1995] E. Automatica. J. Rao and P. Mayne. vol. [Mayne & Michalska 1990] D. Lee. Mayne and H. [Nesi¢ & Laila 2002] D. Control. 47. no. [Michalska & Mayne 1993] H. 36. Rawlings. V. Receding hori- zon control of nonlinear systems. 35. Q. 2000. Receding horizon control and discontinuous state feedback stabilization. H. 2001. Mayne. Morari and J. [Meadows et al. Laila. IEEE Transactions on Automatic Control. vol. 2. S. 38. B. vol. IEEE Transactions on Automatic Control. Automatica. Robust receding horizon control of constrained nonlinear systems. pages 667682. pages 12171229. C. Eaton and J. pages 814824. pages 209217. O. El-Farra and P. Model predictive con- trol: past. pages 50415044. M. vol. 2002. 1999. J. [Morari & H. [Mayne 2001] D. Michalska and D. pages 11531158. 2005. European Journal of Control. Lee 1999] M. 1990. 1993. A. M.Bibliography 243 Type 1 Diabetes. Q. 2006. Constrained model predictive control: Stability and Opti- mality. Control of Constrained Dynamic Systems. Michalska. W. 1995. Q. A Note on Input-to-State Stabilization for Nonlinear Sampled-Data Systems. present and future. [Mhaskar et al. N. Christodes. pages 789814. vol. Rawlings. pages 8799. 4-5. Int. Meadows. Q. Computers and Chemical Engineering. B. vol. H. vol. 62. D. 7. IEEE Transactions on Automatic Control. J. . S. 1. Henson. Mayne. no. Engineering in Medicine and Biology Society. Scokaert. Mhaskar. Robust hybrid predictive control of nonlinear systems.

Systems and Control Letters. 10. 1. vol. R. pages 4960. Teel. pages 16501667. D. pages 259270. 31. Parisini and R. 12. 1999. Nesic and A. 49. 2004. Teel. R. 1999a] D. 1999b] D. 40. vol. Input to State Stability Prop- erties of Networked Control Systems. [Palerm 2003] C. 38. 2004. pages 11031122. no. no. R. . 38. C. KL esimates of discrete time and sampled data nonlinear sys- Systems and Control Letters. Zoppoli. A. Sontag. 1995. vol. A receding-horizon regulator for nonlinear systems and a neural approximation. [Parisini & Zoppoli 1995] T. New York. vol. A framework for stabilization of nonlinear sampled-data systems based on their approximate discrete time models. Nesic and A. Input-Output Stability Prop- erties of Networked Control Systems. Teel and E. Sucient condi- tion for sampled-data stabilization of nonlinear systems via discrete time approximations. Palerm. vol. Nesic and A. Sontag. pages 21212128. Automatica. 49. PhD thesis. no. no. pages 14431451. D. Formulas re- lating tems. 2003. Nesic. Rensselaer Polytechnic Institute. vol. [Nesic & Teel 2004b] D. Automatica. IEEE Transactions on Automatic Control. R. R.244 Bibliography [Nesic & Teel 2004a] D. no. A. IEEE Transactions on Automatic Control. [Nesic & Teel 2004c] D. Nesic. Teel. Troy. 7. Teel and E. [Nesic et al. 4-5. R. Drug Infusion Control: an Extendend Di- rect Model Reference Adaptive Control Strategy. 2004. 1999. [Nesic et al.

L. USA. Sanguinetti and R. Nonlinear stabilization by receding-horizon neural regulators.J. Zoppoli. D. 2008. A. Birkhauser Berlin. Magni. Mexico. L. no. vol. Int. In F. A Model- Based Algorithm for Blood Glucose Control in Type I Diabetic Patients. 2001] M. Parker. editeurs. Ro- bust Receding . 1998] T. M.J. 2. Parisini and D. pages 369 392. pages 148157.A. In American Control Conference. . Badgwell. Yang and K. Washington. June 1113. Ro- bust Model Predictive Control of Nonlinear Systems with Bounded and State-Dependent Uncertainties.S. 2001. Allgower and A. Parisini. pages 16. of the 2001 IEEE International Symposium on Intelligent Control. Seattle. pages 341362. vol. Nonlinear Model Predictive Control. M. pages 16671672. 70. M. Zheng. IEEE Transactions on Biomedical Engineering. Passino. 2008b] G. Parisini. Cooper- ative Search Framework for Distributed Agents. Raimondo. [Qin & Badgwell 2000] S. Pin. 1999. [Pin et al. Provisionally accepted for publication in IEEE Transactions on Automatic Control. In Proc. T. Polycarpou. 1998. Raimondo. 2000. Mexico City. Doyle III and N. F. J. Magni and T. An Overview of Non- linear Model Predictive Control Applications. [Pin et al. Y. Qin and T.Bibliography 245 [Parisini et al. 2008a] G. Control.Horizon Control of Nonlinear Systems with State Dependent Uncertainties: an Input-to-State Stability Approach. Peppas. 1999] R. [Parker et al. Pin. 46. 2008. [Polycarpou et al.

F. 2007. In Proc. Camacho. In IFAC Symp. De- centralized MPC of nonlinear systems: An input-to-state stability approach. A decentralized MPC algorithm for nonlinear systems. France. on Nonlinear Control Systems. Raimondo. Control Engineering Practice. Grenoble. Magni and E. 2007e] D. [Raimondo & Magni 2006] D. 2007b] D. Alamo. Magni and R. on Decision and Control. A. International Journal of Robust and Nonlinear Control. J. 2007. Magni and R. [Raimondo et al. Greece 2-5 July. Raimondo. 2007. [Raimondo et al. on Nonlinear Control Systems. Raimondo and L. Kos. . 2007. 2007. pages 12571262. T. M. M. pages 733764. M. 2007c] D. A survey of industrial model predictive control technology. Badgwell. vol. pages 16511667. L. In IFAC Symp. M. Magni. Camacho. [Raimondo et al. M. D. [Raimondo et al. L. vol. of the 46th IEEE Conf. Scattolini. A robust model predictive control algorithm for nonlinear systems with low computational burden. D. 2003. In IFAC Workshop on Nonlinear Model Predictive Control for Fast Systems. 2007d] D. Limon. 17. Raimondo. Scattolini. Scattolini. Limon and E. 7. In European Control Conference 2007. no. Magni and R. Regional input-to-state stability of min-max model predictive control. M. 2007a] D. De- centralized Open-Loop MPC of Nonlinear Systems: an Input-to-State Stability Approach. Raimondo. [Raimondo et al. Raimondo.246 Bibliography [Qin & Badgwell 2003] S. 11. Qin and T. Towards the practical implementation of min-max non- linear Model Predictive Control. Lazar. M. 2006. L. L. F.

CRC Press. 2006. Magni and E. T. D. no. pages 11361142. IEEE Transactions on Automatic Control. Raimondo. Teel. D. Model-Based Predictive Control: A Practical Approach. F. 2003. 51. A. Mayne and J. O. D. 2008. [Rossiter 2003] JA Rossiter. E. 2000. Simple Robust Control Invariant Tubes for Some Classes of Nonlinear Discrete Time Systems. [Rawlings 2000] JB Rawlings. Submitted to a journal. Q. R. M. 2006b] S. Reachability analysis of discrete-time systems with dis- turbances. Q. Limon and E. L. F. Lygeros. 2008. . Control Systems Magazine. IEEE Transactions on Automatic Control. 3. M. vol. Tutorial overview of model predictive control. IEEE. 2008a] D. [Scokaert & Mayne 1998] P. M. Min-max Feedback Model Predictive Control for Constrained Linear Systems. [Rakovi¢ et al. Raimondo. Mayne. no. Mayne and A. D. V. [Raimondo et al. Alamo. To appear in European Journal of Control. pages 3852. V. 4. on Decision and Control. Camacho.Bibliography 247 [Raimondo et al. Rakovi¢. pages 6397 6402. Rakovi¢. 20. In Proc. Kerrigan. M. of the 45th IEEE Conf. Lazar. Limon. 2008b] D. 2006a] S. Min-Max Nonlinear Model Predictive Control: a Min formulation with guaranteed robust stability. Camacho. 1998. vol. pages 546561. [Rakovi¢ et al. Scokaert and D. Q. Min-max model predictive control of nonlinear systems: a unifying overview on stability. 2006. 43. vol. Astol.

Automatica. 24. [Sontag & Wang 1995b] E. Large-scale dynamic systems. 1997] S. Uehara. Stabilization with respect to noncompact sets: Lyapunov characterizations and eect of bounded inputs. In In Nonlinear Control Systems Design 1992. Shimoda. 1997. T. 1992. M. Y. pages 197211. Sontag and Y.D. 1995. Scokaert. On characterizations of input-to-state stability with respect to compact sets. pages 4349. North-Holland. 1978. 1999. K. AC-44. Mayne and J. Q. [Sontag & Wang 1995a] E. 33. M. Konno. vol. Rawlings. Scokaert. Rawlings and E. Sakakida. On characterizations of the input-to-state stability property.. Sontag and Y. M. Wang. D. Siljak. vol. pages 351359. Shichiri. 8. no. Pergamon Press. [Shimoda et al. New York. 3. Discrete-time Stability with Perturbations: Application to Model Predictive Control. Lin. Frontiers of Medical and Biological Engineering. 1997. S. [Siljak 1978] D. [Scokaert et al. Fliess Ed. Suboptimal Model Predictive Control (Feasibility Implies Stability). 1999] P. pages 463470. M. O. [Sontag & Lin 1992] E. Meadows. Wang.248 Bibliography [Scokaert et al. 1997] P. Closed-loop subcutaneous insulin infusion algorithm with a short-acting insulin analog for long-term clinical application of a wearable articial endocrine pancreas. Oxford. IFAC Symposia Series. D. Nowak and M. vol. D. pages 648 654. J. M. O.D. 1993. B. K. In IFAC Non- . vol. IEEE Transactions on Automatic Control. Sontag and Y. Ichinose. Nishida. B. System & Control Letters.

3. 34.M. Leader to Formation Stability. Darwin. F. New Characterizations of Input-to State-Stability. no. Advanced Drug Delivery Reviews. pages 21292136. on Decision and Control. D. Cambridge. pages 12831294. 41. PhD thesis.Bibliography 249 Linear Control Systems Design Symposium. C. Tahoe City. Sontag and Y. Kumar.the path to physiological glucose control.E. G. Las Vegas. Hariri and M. In Proc. [Tanner et al. Steil. Pantaleon and K. A. J. [Sontag 1989] E. IEEE Transactions on Robotics and Automation. 2006] G. vol. D. Closed-loop insulin delivery . 1995. . Steil. G. pages 24392444. Tanner. [Tanner et al. 2002] H. vol. 1985. vol. [Steil et al. 1996. 1989. IEEE Transactions on Automatic Control. K. pages 226231. 56. Massachussets Institute of Technology (MIT). Sorensen. G. Feasibility of Automating Insulin Delivery for the Treatment of Type 1 Diabetes.M. 2004. (NOLCOS '95). no. 2006. vol. Input-to- state Stability on Formation Graphs. Rebrin. Tanner. Rebrin. Kumar. pages 33443350. of the IEEE Conf. 2004] G. 2004] H. [Sorensen 1985] J. 55. J. 20. Diabetes. G. Pappas and V. A physiological model of glucose metabolism in man and its use to design and assess improved insulin therapies for diabetes. Saad. Smooth stabilization implies coprime factor- ization. [Steil et al.F. Wang. [Sontag & Wang 1996] E. Department of Chemical Engineering. pages 125144. 2002.T. CA. Sontag. Pappas and V. IEEE Transactions on Automatic Control. pages 435443. 2004.

2005] A. IEEE Transactions on Automatic Control. On the Stabilization of Decentralized Control Systems. [Venkat et al. 1. M. Balanc- ing Search and Target Response in Cooperative Unmanned Aerial Vehicle (UAV) Teams. A. [WHO Website ] WHO Website. 44th Conf. 18. 1973. 2007. Man and Cybernetics: B. Wright. pages 571587. vol. no. no. 36. Tessera. Rawlings and S. H. Jin. . IEEE Transactions on Automatic Control. Minai and M. vol. Davidson. vol. Rapport technique. Thesis (in Italian).250 Bibliography [Tessera 2007] R. Raimondo supervisors. 13. S. De Nicolao and D. 5. University of Pavia (Italy). J. Wang and E. 2005. Witsenhausen. pages 473478. Stability and optimality of distributed model predictive control. pages 6680 6685. Venkat. N. [Wang & Davidson 1973] S. IEEE Transactions on Systems. 2006. [Witsenhausen 1968] H. Polycarpou. Sperimentazione in-silico di algoritmi di con- trollo predittivo della glicemia per il diabete di tipo 1. pages 521.who. 2006] Y. L. Jin et al. Liao Y. Seville. 1968. A Minimax Control Problem for Sampled Linear Systems. SP. Decision Contr. http://www. B. Magni G. J. Proc. [Y.int/mediacentre/factsheets/fs312/en/.

- A Brief Tutorial on Recursive Estimation With Examples From Intelligent Vehicle Applications
- 2 Units and Measurement
- SVD-Based Universal Spatial Domain Image Steganalysis
- MIT2 854F10 Stats
- 11.4.2012(2)
- Interpretation Trained Neural Networks Based on Genetic Algorithms
- A First Course on Time Series Analysis
- Ensemble Bma
- Forecast
- ARTICLESPUBL2c Risk Corr Cov Var2
- Development of a Robust Algorithm to Compute Reactive Azeotropes
- Study of Gradient based optimizations for logistic regression
- trupti
- ICICI014P - Widjiantoro, Predictive Control
- Estimation of the Rock Deformation Modulus and RMR
- quality of service management.docx
- Data Pre-processing in the Neural Network Identification
- Potts (2003)_Numerical Analysis, A Virtual Dream or Practical Reality
- Noise Resilient Periodicity Mining In Time Series Data bases
- Jordan Neural Network MPC Maciej
- Singh Thesis
- Comparing 2D Vector Field Visualization Methods- A User Study
- Ethesis Finally
- Elsmar.com_pdf_files_Process Capability - What is Process Capability
- 166_An Introduction to Business Statistics
- Dimensionality Reduction of Markov Chains
- 10.1.1.55.3867
- A Survey of Satellite Imagery Classification with Different Approaches
- Correlations of Continuous Random Data With Major World Events - D.radin,R.nelson

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue listening from where you left off, or restart the preview.

scribd