New trends in optimal filtering and control for polynomial and time-delay systems (Q938604)

The book is concerned with optimal filtering and control problems for some classes of stochastic state space models for which a finite dimensional optimal (or sometimes approximately optimal) solution to such problems exists. In particular, two classes of models are considered. The first is the class of polynomial systems, i.e. stochastic differential systems where the state dynamics or the observation mechanism have drift or diffusion coefficients which are polynomial functions of the state process. The second is the class of linear systems with time delays affecting the state or observation process. Besides filtering and control the book deals with other topics connected to these models, namely parameter estimation and the problem of robustification of control and filtering algorithms. The proposed optimal (or approximately optimal) finite dimensional solutions are also applied to specific examples and compared with other approximate solutions obtained mainly using linearized models or models without delays. Only in a few instances examples connected to concrete applicative problems are discussed. The presentation of the subjects treated in the book is very detailed. In fact, besides the general cases, also many particular situations are discussed, with continuous repetitions of the structure of the various sections and even of the sentences illustrating the steps of the arguments involved. A detailed description of the content of the book and, when useful, a few comments will be given below. To this end some remarks will be useful to clarify some terminology adopted. First, notice that by ``completely measured states'' the author means that the observation and state processes have the same dimension, the observation depends linearly on the state and the corresponding read-out matrix is invertible. Second, by ``partially measured states'' it is meant that the dimension of the observation process is less than the dimension of the state. On the other hand, by ``completely observed states'' it is meant, as usual, that the states are perfectly known. In the present review the same terminology will be used. Chapter 1 is devoted to optimal nonlinear filtering for polynomial systems. In particular Section 1.1 is concerned with the case of linear observations with completely measured states. Models used here are characterized by a state dynamics with a drift which is a polynomial function of the state while the diffusion coefficient does not depend on the state. Examples of practical importance are also examined, namely filters for a third order automotive system and for a bilinear terpolymerization process. Section 1.2 deals with polynomial systems with a partially measured linear part, i.e. the models considered have some components of the state which are completely measured while the other components have a linear dynamics but are only partially measured. In Section 1.3 polynomial state models are considered with multiplicative noise. In the author's terminology this means that both the drift and the diffusion coefficients are polynomial functions of the state process. In Section 1.4 polynomial state models with both a partially measured linear part and multiplicative noise are examined. In this framework the cubic sensor problem, namely a model with linear state dynamics and third-order read-out map, is studied using an extended state space, where the added components consist of the third powers of the original states. The initial conditions of these added components are assumed independently of the initial conditions of the original state components. The author asserts that the resulting filter is optimal, but it is clear that it is only some sort of of approximate solution to the optimal filtering problem. A generalization of the cubic sensor problem considers systems with linear state and polynomial observations and is studied in Section 1.5 where the problem is reduced to the case of polynomial state with a partially measured linear part and multiplicative noise treated in Section 1.4. The proposed filters are again some sort of approximate filters, but they are considered as optimal by the author. In Chapter 2 optimal identification and control problems are examined. Section 2.1 deals with joint filtering and parameter estimation for linear completely measured systems, where the drift coefficient in the state dynamics is a polynomial function of unknown parameters. Such parameters are modelled as Wiener processes and, using an extended state space, the problem is reduced to a filtering problem for polynomial systems with a partially measured linear part and linear observations (treated in Section 1.2). In Section 2.2 dual optimal control problems for deterministic polynomial systems are considered. Here the state is completely observed and a quadratic criterion is adopted. A duality principle between the control problem and the filtering problem for polynomial systems with linear observations (examined in Section 1.1) is conjectured and this produces a control strategy whose optimality is then proved. Applications to concrete situations include optimal regulators for a terpolymerization reactor and for an automotive system. In Section 2.3 optimal controllers for polynomial systems with partial observations and quadratic criterion are studied. To this end a separation principle is proved and the optimal controller is then derived and applied to an automotive system. In Chapter 3 the problem of optimal filtering for linear time-delay systems is examined. Here, except for Section 3.7, only linear state models are considered, so that partially measured states are allowed. For the derivation of the filters the author uses some ``conditional expectation equalities'' which are frequently obscure and possibly incorrect. The filters obtained in this way are in some sense approximate filters and their behaviour is compared to that of Kalman-Bucy filters corresponding to the systems without delays. Various particular situations are considered: state dynamics without delays and observations with multiple delays (Section 3.1), state dynamics with a single delay and observations without delays (Sections 3.2 and 3.6), both state dynamics and observations with a single delay (Section 3.3), state dynamics with a single delay and observations with multiple delays (Section 3.4), both state dynamics and observations with multiple delays (Section 3.5). Finally, in Section 3.7 the case of polynomial state dynamics and observations with a single delay is studied. Chapter 4 deals with deterministic optimal control problems for linear systems with time delays. The technique used parallels the one in Section 2.2, namely a duality principle between the control problem and the filtering problem for linear systems with delays in the observations (examined in Chapter 3) is conjectured and this provides a candidate control strategy whose optimality is then proved. The behaviour of such strategies is then compared to that of the corresponding linear regulators without delays and of other linear regulators based on rational approximations of the input-state transfer function. However, such rational approximations often result in computational difficulties and unsatisfactory behaviours. Various special cases are then considered in detail: linear systems with multiple input delays (Section 4.1), linear systems with an equal single state and input delay (Section 4.2), linear systems with multiple state delays (Section 4.3), linear systems with multiple state and input delays (Section 4.4). Finally, in Section 4.5 an optimal controller for partially observable linear systems with single input and observation delays is studied. Such controller is based on a separation principle and uses the results in Section 4.1 (control for linear systems with multiple input delays) and in Section 3.1 (filtering over observations with multiple delays). In the final Chapter 5 sliding mode applications to optimal filtering and control are discussed. Here the state and observation models are influenced by uncertain inputs modelling nonlinearities and (possibly smooth) disturbances. The aim is to robustify the filter and/or the controller with the aim of eliminating the uncertain disturbances (filtering problem) and/or following the nominal state trajectory corresponding to the state dynamics without the uncertain inputs (control problem). In Section 5.1 linear systems with a single input delay in the state dynamics, observations with a single state delay and uncertain disturbances depending on the current and delayed state are considered. The problem is to robustify the optimal controller of Section 4.5 for partially observable linear systems with input and observation delays. The sliding mode integral technique provides in this situation a joint compensation of state and observation uncertain inputs. The case of a nonlinear state drift is also examined. In Section 5.2 a completely observable linear system with a single state delay is studied and an integral sliding mode controller is proposed with the aim of following the nominal trajectory corresponding to the undisturbed system. The robust control of an analogous system with a nonlinear drift is also considered.