Linear quadratic Pareto optimal control problem of stochastic singular systems

Linear quadratic Pareto optimal control problem of stochastic singular systems

Author’s Accepted Manuscript Linear quadratic pareto optimal control problem of stochastic singular systems Weihai Zhang, Yaning Lin, Lingrong Xue ww...

642KB Sizes 0 Downloads 45 Views

Author’s Accepted Manuscript Linear quadratic pareto optimal control problem of stochastic singular systems Weihai Zhang, Yaning Lin, Lingrong Xue

www.elsevier.com/locate/jfranklin

PII: DOI: Reference:

S0016-0032(16)30440-9 http://dx.doi.org/10.1016/j.jfranklin.2016.11.021 FI2810

To appear in: Journal of the Franklin Institute Received date: 22 January 2016 Revised date: 25 July 2016 Accepted date: 21 November 2016 Cite this article as: Weihai Zhang, Yaning Lin and Lingrong Xue, Linear quadratic pareto optimal control problem of stochastic singular systems, Journal of the Franklin Institute, http://dx.doi.org/10.1016/j.jfranklin.2016.11.021 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting galley proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Linear Quadratic Pareto Optimal Control Problem of Stochastic Singular Systems



Weihai Zhang †, Yaning Lin , Lingrong Xue College of Electrical Engineering and Automation, Shandong University of Science and Technology, Qingdao 266590, Shandong Province, P. R. China

November 24, 2016

1. College of Electrical Engineering and Automation, Shandong University of Science and Technology, Qingdao 266590, China; 2. College of Science, Shandong University of Technology, Zibo 255000, China.

Abstract: This paper is concerned with the finite horizon linear quadratic (LQ) Pareto optimal control problem of stochastic singular systems. By means of the square completion technique, for the finite horizon LQ optimal control of stochastic singular systems, we establish a new kind of generalized differential Riccati equations (GDREs) and present the existence condition of the solution of the GDREs. Then, for the finite horizon LQ Pareto optimal control, it is shown that under the solvability of the corresponding GDREs, all Pareto candidates can be obtained by solving a weighting sum optimal control. Finally, an example is provided to show the effectiveness of our main results. Keywords: Pareto optimality; stochastic singular systems; finite horizon LQ optimal control; generalized differential Riccati equations. ∗ This

research was supported by NSF of China (Nos.61573227, 61633014), the Research Fund for the Taishan

Scholar Project of Shandong Province of China, the SDUST Research Fund (No.2015TDJH105) and the State Key Laboratory of Alternate Electrical Power System with Renewable Energy Sources(Grant No. LAPS16011) † Corresponding

author. E-mail: w [email protected] (W.Zhang).

1

1. Introduction Singular systems, which are also referred to as descriptor systems, generalized state-space systems, differential algebraic systems, or implicit systems, have extensive applications in many practical systems, for example, circuit boundary control systems, chemical processes, electrical networks, economy systems and other areas [7]. Hence a great number of fundamental notions and theories for singular systems have been researched, such as stability and stabilization [29, 31, 38], LQ optimal control [28, 30, 37], H∞ control [19, 26] and so on. As it is well known, environmental noise exists and cannot be neglected in many dynamical systems [5, 14, 20, 25]. So, singular systems with stochastic noise are more realistic mathematical models; see Example 5.2 in [33] and system (5) in [36]. There have been some reports on the study of stochastic singular systems, but most of which are concerned with the stability and stabilization. In [5], the author developed stochastic stability and stochastic stabilization criteria for stochastic singular systems, and designed two kinds of state feedback controllers to stochastically stabilize the systems. In [13], the authors gave a complete proof of mean-square exponential stability for singular stochastic Itˆo systems, and then investigated observer-based controller designs. The reference [35] discussed stability criteria for stochastic singular systems with state-dependent noise in both continuous-time and discrete-time cases. In [27], the authors studied the stochastic stabilization problem for a class of singular Markovian jump systems, and proposed a new kind of stochastic controller design such that the closed-loop system has a unique solution and is almost surely exponentially admissible. However, there are few articles talking about the optimal control problem of stochastic singular systems. In [33], under some assumptions, the authors developed a sufficient condition for mean-square admissibility in terms of LMI, investigated finite-time horizon and infinite-time horizon LQ control problems using square completion technique, and discussed some results involving new stochastic generalized Riccati equation. The dynamic games of continuous-time and discrete-time systems have been extensively investigated by many researchers (see, e.g. [3] and the reference therein). Among various dynamic games, the Pareto game arises when multiple players, affecting a dynamic system, coordinate their actions/controls in order to optimize their objectives, which has been widely used

2

in economic theory such as optimal economic growth, environmental economics and engineering [1, 3, 8, 9, 10, 22, 23, 24]. The set of Pareto efficient equilibria for the indefinite cooperative LQ control problem of linear affine systems was determined in [9]. Necessary and sufficient conditions for the existence of a Pareto optimum of the finite horizon cooperative differential games of nonlinear systems were presented in [10]. The reference [23] derived conditions for the existence of Pareto optimal solutions for the LQ infinite horizon cooperative differential games. In [24], the necessary and sufficient conditions for the existence of Pareto solutions in infinite horizon cooperative differential games with open loop information structure were discussed. However, most of the reports are about the deterministic systems. To the best of our knowledge, the LQ pareto optimal control of stochastic singular systems is still unsolved. In this paper, we concentrate our attention on the finite horizon LQ Pareto optimality problem for a class of continuous-time stochastic Itˆ o singular systems. A well known way to obtain Pareto optimal controls is to solve a weighting sum optimal control problem [16, 32]. Every control minimizing a weighting sum of the cost functionals of all players (where all weights are in the unit simplex) is Pareto optimal. So, we first discuss the general finite horizon LQ optimal control problem in Section 2. Using the square completion technique, we give a sufficient condition for the well-posedness by introducing a new type of GDREs. Furthermore, we put forward the existence condition of the solution of the GDREs. Next, in Section 3, we consider the finite horizon stochastic singular LQ Pareto optimal control problem. Via the discussion of convexity of the cost functional, under some assumptions, it is revealed that the solvability of the GDREs gives a sufficient condition under which we can find all Pareto efficient solutions by the weighting sum optimization method. Finally, an illustrative example is given to show the effectiveness of the proposed results. Notations:

Rn : the set of all real n-dimensional vectors. Rm×n : the set of all m × n real

matrices. A > 0 (resp. A ≥ 0): A is a real symmetric positive definite (resp. positive semidefinite) matrix. In : n-dimensional identity matrix. AT : the transpose of matrix A. E(x): the mathematical expectation of x. tr(A): the trace of a square matrix A. A := {α = (α1 , α2 )|αi ≥ 0 and

2

i=1

αi = 1}. L2 [0, T ] := {u(t)|u(t) is a square integrable function on [0, T ]}. C 1,2 (R+ × Rn ):

the class of functions V (t, x) once continuously differential with respect to t ∈ R+ and twice continuously differential with respect to x ∈ Rn , except possibly at the point x = 0. 3

2. Finite Horizon Singular LQ Optimal Control In this section, we consider the following stochastic singular system ⎧   ⎪ ⎪ ⎨ Edx(t) = Ax(t) + Bu(t) dt + A x(t)dw(t), p

⎪ ⎪ ⎩ Ex(0) = x0 ,

(1)

where x(t) ∈ Rn is the system state vector, u(t) ∈ Rm is the control vector, w(t) is a onedimensional standard Wiener process that is defined on the complete probability space (Ω, F , P). E, A, B, Ap are known matrices of appropriate dimensions and rank(E) = r ≤ n; x0 ∈ Rn is the compatible initial condition which is deterministic. Define the set of all admissible controls Uad = {u(·) ∈ L2 [0, T ]| the corresponding solution x(·) of (1) satisfies x(·) ∈ L2 [0, T ]}. For each x0 and u(·) ∈ Uad , we consider the following cost functional J(0, T, u, x0 )  E



0

T

 T  x (t)Qx(t) + uT (t)Ru(t) dt + xT (T )E T QT Ex(T ) ,

(2)

where Q = QT , QT = QTT and R > 0. The objective of the optimal control problem in this section is to find the optimal control u∗ (t) ∈ Uad that minimizes the cost functional J(0, T, u, x0 ). The optimal value function is defined as V (0, T, x0 )  min J(0, T, u, x0 ). u∈Uad

Definition 2.1. The optimization problem (1)-(2) is called well-posed if V (0, T, x0 ) > −∞. u∗ (t) is said to be optimal if u∗ (t) achieves the minimum of J(0, T, u, x0 ). Next, we give some assumptions which are necessary for the proofs of our main results. Assumption 2.1. rank(E ⎡ Ap ) = rank(E). ⎤ ⎢ E Assumption 2.2. rank ⎢ ⎣ 0

A E

B ⎥ ⎥ = n + rankE. ⎦ 0

Remark 2.1. Assumption 2.2 is a basic requirement for singular controlled systems [37]. Under Assumption 2.2, we can choose an appropriate control to eliminate the impulse phenomena

4

in system (1). Furthermore, Assumption 2.1 and Assumption 2.2 can guarantee the existence and uniqueness of the impulse-free solution of the stochastic singular system (1). When B = 0, Assumption 2.1 and Assumption 2.2 reduce to Assumption 1 in [13]. On this point, we have the following conclusion. Lemma 2.1. If Assumption 2.1 and Assumption 2.2 hold, then system (1) has a unique solution on [0, +∞), in which there is no impulse. Proof. Considering the singular value decomposition (SVD) of matrix E, under Assumption 2.1, there exist two orthogonal matrices M, N ∈ Rn×n such that ⎡ ⎤ ⎡ ⎢ Σr M EN = ⎢ ⎣ 0

0 ⎥ ⎥, ⎦ 0

⎢ Ap1 M Ap N = ⎢ ⎣ 0



Ap2 ⎥ ⎥, ⎦ 0

(3)

where Σr ∈ Rr×r is a nonsingular diagonal matrix and Ap1 ∈ Rr×r , Ap2 ∈ Rr×(n−r) . Accordingly, define





⎢ A11 M AN = ⎢ ⎣ A21





A12 ⎥ ⎥, ⎦ A22

⎢ B1 ⎥ ⎥, MB = ⎢ ⎦ ⎣ B2

(4)

where A11 ∈ Rr×r , A12 ∈ Rr×(n−r) , A21 ∈ R(n−r)×r , A22 ∈ R(n−r)×(n−r) , B1 ∈ Rr×m , B2 ∈ R(n−r)×m and let





⎢ ξ1 (t) ⎥ ⎥, ξ(t) = N −1 x(t) = ⎢ ⎣ ⎦ ξ2 (t)

(5)

where ξ1 (t) ∈ Rr , ξ2 (t) ∈ Rn−r . By (3)-(5), system (1) can be transformed into ⎧ ⎪ ⎪ ⎪ Σr dξ1 (t) = [A11 ξ1 (t) + A12 ξ2 (t) + B1 u(t)]dt ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ + [Ap1 ξ1 (t) + Ap2 ξ2 (t)]dw(t), ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩

(6)

ξ1 (0) = [Σ−1 r 0]M x0 , 0 = A21 ξ1 (t) + A22 ξ2 (t) + B2 u(t).

 By Assumption 2.2, the matrix A22

 B2 has full row rank, so there exists a nonsingular matrix

X ∈ R(n−r+m)×(n−r+m) such that 

 A22

B2

 X=

5

 In−r

0

.

Denoting





⎢ X11 X ⎢ ⎣ X21

X12 ⎥ ⎥, ⎦ X22





B1 ⎥ ⎢ A¯12 ⎥X  ⎢ ⎦ ⎣ A¯p2 0

⎢ A12 ⎢ ⎣ Ap2





¯1 ⎥ B ⎥, ⎦ ¯ D

where X11 ∈ R(n−r)×(n−r) , X12 ∈ R(n−r)×m , X21 ∈ Rm×(n−r) , X22 ∈ Rm×m , A¯12 ∈ Rr×(n−r) , ¯1 ∈ Rr×m , A¯p2 ∈ R(n−r)×(n−r) , D ¯ ∈ R(n−r)×m . Taking the nonsingular transformation B ⎡ ⎤⎡ ⎤ ⎡ ⎤ ⎢ ξ1 (t) ⎥ ⎢ Ir ⎢ ⎥ ⎢ ⎢ ⎥ ⎢ ⎢ ξ (t) ⎥ = ⎢ 0 ⎢ 2 ⎥ ⎢ ⎢ ⎥ ⎢ ⎣ ⎦ ⎣ 0 u(t)

0

0

X11

X12

X21

X22

⎥ ⎢ ξ1 (t) ⎥ ⎥⎢ ⎥ ⎥⎢ ⎥ ⎥ ⎢ ξ¯ (t) ⎥ , ⎥⎢ 2 ⎥ ⎥⎢ ⎥ ⎦⎣ ⎦ u ¯(t)

where ξ¯2 (t) ∈ Rn−r , u¯(t) ∈ Rm . By (7), system (6) is transformed as ⎧   ⎪ ⎪ ¯1 u ⎪ Σr dξ1 (t) = (A11 − A¯12 A21 )ξ1 (t) + B ¯(t) dt ⎪ ⎪ ⎪ ⎪ ⎪   ⎪ ⎪ ¯ u¯(t) dw(t), ⎨ + (Ap1 − A¯p2 A21 )ξ1 (t) + D ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩

(7)

(8)

ξ1 (0) = [Σ−1 r 0]M x0 , ξ¯2 (t) = −A21 ξ1 (t).

The first equation of (8) is an ordinary stochastic differential equation, in which ξ1 (t) is the state vector and u ¯(t) is the control vector. According to Theorem 3.1 in [21], the first equation of (8) has 0]M x0 . Accordingly, a unique solution ξ1 (t) on [0, +∞) under the initial condition ξ1 (0) = [Σ−1 r ξ¯2 (t) = −A21 ξ1 (t) exists and is unique. Thus, system (8) has a unique impulse-free solution on [0, +∞), which implies that system (1) has a unique impulse-free solution on [0, +∞). Now, we are in a position to give the main results of the LQ optimal control problem of stochastic singular systems. First, we give the following useful lemmas.   Lemma 2.2. [15] Let V t, x(t) ∈ C 1,2 (R+ × Rn ), x(t) satisfies the following stochastic differential equation dx(t) = f (t)dt + g(t)dw(t),   then V t, x(t) is an Itˆ o’s process and       dV t, x(t) = LV t, x(t) dt + Vx t, x(t) g(t)dw(t), where          1  LV t, x(t) = Vt t, x(t) + Vx t, x(t) f (t) + tr g T (t)Vxx t, x(t) g(t) . 2 6

Lemma 2.3. [2] Let a matrix M ∈ Rm×n be given, then there exists a unique matrix M + ∈ Rn×m , which is called the Moore-Penrose pseudo inverse of M , such that ⎧ ⎪ ⎪ ⎨ M M + M = M, M + M M + = M + , ⎪ ⎪ ⎩

(M M + )T = M M + , (M + M )T = M + M.

The following GDREs for matrix P (t) play an important role in the study of the LQ optimal control of stochastic singular system ⎧ −E T P˙ (t) = AT P (t) + P T (t)A + Q + ATp (E + )T P T (t)EE + Ap ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ − P T (t)BR−1 B T P (t), ⎨

(9a)

⎪ ⎪ ⎪ E T P (T ) = E T QT E, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ E T P (t) = P T (t)E.

(9b) (9c)

Remark 2.2. (i)When E = I, system (1) degenerates to the deterministic stochastic system. (9) can be regarded as an extension of the GDRE in [34]. (ii) When Ap = 0, system (1) reduces to the deterministic singular system. In this case, (9) is consistent with the differential Riccati equations (DREs) of [30]. Make a transformation for P (t),





⎢ P11 (t) M −T P (t)N = ⎢ ⎣ P21 (t)

P12 (t) ⎥ ⎥, ⎦ P22 (t)

(10)

where P11 (t) ∈ Rr×r , P12 (t) ∈ Rr×(n−r) , P21 (t) ∈ R(n−r)×r , P22 (t) ∈ R(n−r)×(n−r) . Accordingly, denote







⎢ Q11 ˆ  N T QN = ⎢ Q ⎣ QT12

Q12 ⎥ ⎥, ⎦ Q22



⎢ QT 11 M −T QT M −1 = ⎢ ⎣ QTT 12

QT 12 ⎥ ⎥, ⎦ QT 22

(11)

where Q11 , QT 11 ∈ Rr×r , Q12 , QT 12 ∈ Rr×(n−r) , Q22 , QT 22 ∈ R(n−r)×(n−r) . In order to discuss the existence condition of the solution for the GDREs (9), we need the following assumptions. Assumption 2.3. Q ≥ 0 and QT ≥ 0. ˆ ≥ 0 and QT 11 ≥ 0, so there exist two matrices C1 ∈ Rn×r and Under Assumption 2.3, Q C2 ∈ Rn×(n−r) such that

⎡ ⎢ Q11 ˆ=⎢ Q ⎣ QT12







 Q12 ⎥ ⎢ ⎥ ⎥=⎢ ⎥ C 1 ⎦ ⎣ ⎦ Q22 C2T C1T

7

 C2

.

 Assumption 2.4. The triple A22

B2

 C2 is completely controllable and observable.

Assumption 2.5. λi + λj = 0, where λi and λj are arbitrary eigenvalues of A22 . On the existence condition of the solution of GDREs (9), we have the following theorem. Theorem 2.1. If one of the following two conditions is satisfied: (1) Under Assumption 2.3 and Assumption 2.4, Ap2 = 0; (2) Under Assumption 2.3 and Assumption 2.5, B2 = 0, then the GDREs (9) has a solution P (t) with Σr P11 (t) ≥ 0 and P12 (t) = 0. Proof. Using the relation (3), the Moore-Penrose pseudo inverse of E is ⎡ ⎤ ⎢ Σ−1 r E+ = N ⎢ ⎣ 0

0 ⎥ ⎥ M. ⎦ 0

By relations (3), (4), (10) and (11), equation (9a) can be partitioned into ⎤ ⎡ ⎤⎡ ˙ ˙ ⎢ Σr 0 ⎥ ⎢ P11 (t) P12 (t) ⎥ ⎥ ⎥⎢ − ⎢ ⎦ ⎣ ⎦⎣ ˙ ˙ P21 (t) P22 (t) 0 0 ⎡ ⎤⎡ ⎤ ⎢ AT11 = ⎢ ⎣ AT12 ⎡

AT21 ⎥ ⎢ P11 (t) ⎥⎢ ⎦⎣ T P21 (t) A22 ⎤⎡

⎢ +⎢ ⎣

T (t) P11

⎢ +⎢ ⎣

ATp1

P12 (t) ⎥ ⎥ ⎦ P22 (t)

T P21 (t)

⎥ ⎢ A11 ⎥⎢ ⎦⎣ T T A21 P12 (t) P22 (t) ⎡ ⎤⎡ ⎤⎡



ATp2

0 ⎥ ⎢ Σr ⎥⎢ ⎦⎣ 0 0

T (t) ⎢ P11 −⎢ ⎣ T (t) P12







A12 ⎥ ⎢ Q11 Q12 ⎥ ⎥+⎢ ⎥ ⎦ ⎣ ⎦ T A22 Q12 Q22 ⎤⎡ ⎤⎡ T (t) P11

T P21 (t)

0 ⎥⎢ ⎥ ⎢ Σr ⎥⎢ ⎥⎢ ⎦⎣ ⎦⎣ T T 0 P12 (t) P22 (t) 0 ⎤⎡ ⎤⎡

T P21 (t) ⎥ ⎢ S11 ⎥⎢ ⎦⎣ T T P22 (t) S12

S12 ⎥ ⎢ P11 (t) ⎥⎢ ⎦⎣ S22 P21 (t)

0 ⎥⎢ ⎥⎢ ⎦⎣ 0 ⎤

(12) ⎤⎡ Σ−1 r 0

0 ⎥ ⎢ Ap1 ⎥⎢ ⎦⎣ 0 0

⎤ Ap2 ⎥ ⎥ ⎦ 0

P12 (t) ⎥ ⎥ ⎦ P22 (t)

with the boundary condition Σr P11 (T ) = Σr QT 11 Σr ≥ 0,

(13)

where S11 = B1 R−1 B1T , S12 = B1 R−1 B2T , S22 = B2 R−1 B2T . By (9c), we get Σr P11 (t) =

8

(Σr P11 (t))T and P12 (t) = 0. Then, from (12), we obtain three equations as follows: ⎧ T −1 T T ⎪ (Σr P˙11 (t)) = −Q11 − (Σr P11 (t))T (Σ−1 r A11 ) − P21 (t)A21 − (Σr A11 ) (Σr P11 (t)) − A21 P21 (t) ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ T −1 T T ⎪ ⎪ + (Σr P11 (t))T (Σ−1 ⎪ r S12 )P21 (t) + P21 (t)(Σr S12 ) (Σr P11 (t)) + P21 (t)S22 P21 (t) ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ −1 −1 T −1 ⎪ ⎪ + (Σr P11 (t))T (Σ−1 ⎪ r S11 Σr )(Σr P11 (t)) − (Σr Ap1 ) (Σr P11 (t))(Σr Ap1 ), (14a) ⎪ ⎪ ⎪ ⎨ T T T T (t)A21 − (Σ−1 0 = −QT12 − P22 r A12 ) (Σr P11 (t)) − A22 P21 (t) + P22 (t)S22 P21 (t) ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ T T −1 T −1 ⎪ ⎪ (t)(Σ−1 (14b) + P22 r S12 ) (Σr P11 (t)) − (Σr Ap2 ) (Σr P11 (t))(Σr Ap1 ), ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ T T ⎪ ⎪ (t)A22 − AT22 P22 (t) + P22 (t)S22 P22 (t) 0 = −Q22 − P22 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ T −1 (14c) − (Σ−1 r Ap2 ) (Σr P11 (t))(Σr Ap2 ). (1) If Ap2 = 0, (14) can be rewritten as ⎧ T −1 T T (Σr P˙11 (t)) = −Q11 − (Σr P11 (t))T (Σ−1 ⎪ r A11 ) − P21 (t)A21 − (Σr A11 ) (Σr P11 (t)) − A21 P21 (t) ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ T −1 T T ⎪ + (Σr P11 (t))T (Σ−1 ⎪ r S12 )P21 (t) + P21 (t)(Σr S12 ) (Σr P11 (t)) + P21 (t)S22 P21 (t) ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ −1 −1 T −1 ⎪ + (Σr P11 (t))T (Σ−1 ⎨ r S11 Σr )(Σr P11 (t)) − (Σr Ap1 ) (Σr P11 (t))(Σr Ap1 ), (15a) ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩

T T T T (t)A21 − (Σ−1 0 = −QT12 − P22 r A12 ) (Σr P11 (t)) − A22 P21 (t) + P22 (t)S22 P21 (t) T T (t)(Σ−1 + P22 r S12 ) (Σr P11 (t)),

(15b)

T T (t)A22 − AT22 P22 (t) + P22 (t)S22 P22 (t). 0 = −Q22 − P22

(15c)

(15c) is an algebraic Riccati equation. Considering Q22 = C2T C2 ≥ 0 and S22 = B2 R−1 B2T with R > 0, under Assumption 2.4, (15c) has a unique positive definite solution P22 (t). Substituting P22 (t) into (15b), we get P21 (t) = L2 (t) + L1 (t)(Σr P11 (t)), where  −T  −1 T (Σr A12 ) − (Σ−1 L1 (t) = − A22 − S22 P22 (t) r S12 )P22 (t) ,  −T  T Q12 + AT21 P22 (t) . L2 (t) = − A22 − S22 P22 (t) Furthermore, substituting P21 (t) into (15a), and making some calculations, we get (Σr P˙11 (t)) = −Q0 − (Σr P11 (t))T A0 − AT0 (Σr P11 (t)) + (Σr P11 (t))T S0 (Σr P11 (t)) T −1 − (Σ−1 r Ap1 ) (Σr P11 (t))(Σr Ap1 )

9

(16)

with the boundary condition (13), where Q0 = Q11 + LT2 (t)A21 + AT21 L2 (t) − LT2 (t)S22 L2 (t), T −1 T A0 = Σ−1 r A11 + L1 (t)A21 − (Σr S12 )L2 (t) − L1 (t)S22 L2 (t), T −1 T −1 T S0 = Σ−1 r S11 + L1 (t)(Σr S12 ) + (Σr S12 )L1 (t) + L1 (t)S22 L1 (t).

Similar to the proof of Theorem 1 in [28], we can prove that there exist B0 and C0 such that S0 = B0 R−1 B0T , Q0 = C0T C0 . By [4], there exists a unique symmetric positive semi-definite solution Σr P11 (t) ≥ 0 to the DRE (16). Therefore, ⎡ ⎢ P11 (t) P (t) = M T ⎢ ⎣ P21 (t)

⎤ P12 (t) ⎥ ⎥ N −1 ⎦ P22 (t)

is the solution of (9), which satisfies Σr P11 (t) ≥ 0 and P12 (t) = 0. (2) If B2 = 0, then S12 = S22 = 0.   and Considering P12 (t) = 0, pre-multiplying and post-multiplying (12) by Ir − AT21 A−T 22 

Ir − AT21 A−T 22

T

respectively, we obtain

  −1 −1 (Σr P˙11 (t)) = −(Σr P11 (t))T (Σ−1 r A11 ) − (Σr A12 )A22 A21  T −1 −1 − (Σ−1 (Σr P11 (t)) r A11 ) − (Σr A12 )A22 A21     −T T T −1 ˆ + (Σr P11 (t))T (Σ−1 − Ir − AT21 A−T r S11 Σr )(Σr P11 (t)) 22 Q Ir − A21 A22

(17)

T    −1 −1 −1 −1 (Σr P11 (t)) (Σ−1 − (Σ−1 r Ap1 ) − (Σr Ap2 )A22 A21 r Ap1 ) − (Σr Ap2 )A22 A21 ˆ ≥ 0, and S11 = B1 R−1 B T with R > 0, according with the boundary condition (13). Considering Q 1 to [4], the DRE (17) has a unique symmetric positive semi-definite solution Σr P11 (t) ≥ 0 on [0, T ]. Substituting Σr P11 (t) and S22 = 0 into (14c), and considering Q22 = C2T C2 , we have ⎡ ⎤T ⎡ ⎤ ⎢ T AT22 P22 (t) + P22 (t)A22 + ⎢ ⎣

C2 1

(Σr P11 (t)) 2 (Σ−1 r Ap2 )

⎥ ⎢ ⎥ ⎢ ⎦ ⎣

C2 1

(Σr P11 (t)) 2 (Σ−1 r Ap2 )

⎥ ⎥ = 0. ⎦

(18)

Under Assumption 2.5, (18) has a unique solution P22 (t). Substituting Σr P11 (t) and P22 (t) into (14b), and considering S12 = S22 = 0, we get   T T −1 T −1 − QT12 − P22 P21 (t) = A−T (t)A21 − (Σ−1 r A12 ) (Σr P11 (t)) − (Σr Ap2 ) (Σr P11 (t))(Σr Ap1 ) . 22 Therefore, the GDREs (9) have a solution ⎡



⎢ P11 (t) P12 (t) ⎥ −1 ⎥N , P (t) = M T ⎢ ⎦ ⎣ P21 (t) P22 (t) 10

in which Σr P11 (t) ≥ 0 and P12 (t) = 0. The conclusion is proved. Remark 2.3. We give a sufficient but not necessary condition for the existence of the solution to GDREs (9). That is to say, even if the conditions in Theorem 2.1 are not established, the solution to GDREs (9) may also exist. Please refer to the example in Section 4. Remark 2.4. When Ap1 = Ap2 = 0, i.e. Ap = 0, GDREs (9) degenerates to DREs of deterministic singular systems of [30]. In this case, Assumption 2.3 and Assumption 2.4 are consistent with the existence conditions of the solution in [30]. For this case, our results extended the existing results in [30]. Based on the condition for the unique solvability of Sylvester equation [12], Assumption 2.5 can guarantee the existence and uniqueness of the solution of (18). For the well-posedness of the finite horizon LQ optimal control problem (1)-(2), we have the following conclusion. Theorem 2.2. If the GDREs (9) admit a solution on [0, T ], then the finite horizon LQ optimal control problem (1)-(2) admits uncountably many feedback optimal control strategies given by u∗ (t) = −R−1 B T K(t)x(t), where

(19) ⎤

⎡ ⎢ K(t) = M T ⎢ ⎣ ⎡

P11 (t)

0

P21 (t) + P22 (t)L(t) − F (t)L(t) F (t) ⎤

⎥ −1 ⎥N ⎦

⎢ P11 (t) P12 (t) ⎥ −1    ⎥ N being a solution of (9), L(t) = − A22 −S22 P22 (t) −1 A21 − with P (t) = M T ⎢ ⎦ ⎣ P21 (t) P22 (t)  T S12 P11 (t) − S22 P21 (t) and F (t) is any time-varying matrix making A22 − S22 F (t) invertible within the interval [0, T ]. Additionally, the minimum value of the cost functional (2) is given by J(0, T, u∗ , x0 ) = xT (0)E T P (0)x(0). Proof. Under the condition E T P (t) = P T (t)E, applying Itˆo’s formula, we obtain      T   d xT (t)E T P (t)x(t) = d Ex(t) P (t)x(t) + xT (t)E T dP (t) x(t) + xT (t)P T (t)d Ex(t)    T + d Ex(t) (E + )T P T (t)EE + d Ex(t)  T = Ax(t) + Bu(t) P (t)x(t) + xT (t)E T P˙ (t)x(t)   + xT (t)P T (t) Ax(t) + Bu(t)

+ xT (t)ATp (E + )T P T (t)EE + Ap x(t) dt + · · · dw(t), 11

(20)

where



···



does not affect the calculation results and can be omitted.

Integrating (20) from 0 to T , taking expectations on the both sides, one gets   E xT (T )E T P (T )x(T ) − xT (0)E T P (0)x(0) T  T =E Ax(t) + Bu(t) P (t)x(t) + xT (t)E T P˙ (t)x(t)

(21)

0

   + xT (t)P T (t) Ax(t) + Bu(t) + xT (t)ATp (E + )T P T (t)EE + Ap x(t) dt . Adding (21) to (2), and using the square completion technique, we have J(0, T, u, x0 ) T  xT (t) Q + AT P (t) + P T (t)A + E T P˙ (t) =E 0

+

ATp (E + )T P T (t)EE + Ap T  −1 T

+E

0

u(t) + R

 − P T (t)BR−1 B T P (t) x(t)dt

T   B P (t)x(t) R u(t) + R−1 B T P (t)x(t) dt

   + E x (T )E T QT Ex(T ) − E xT (T )E T P (T )x(T ) + xT (0)E T P (0)x(0) 

T

= xT (0)E T P (0)x(0) T  T   +E u(t) + R−1 B T P (t)x(t) R u(t) + R−1 B T P (t)x(t) dt . 0

Obviously, when u(t) = −R−1 B T P (t)x(t),

(22)

J(0, T, u, x0 ) attains its minimum xT (0)E T P (0)x(0). Now, substituting (22) into (6), we obtain   ⎧ Σr dξ1 (t) = A11 ξ1 (t) + A12 ξ2 (t) − S11 P11 (t)ξ1 (t) − S12 P21 (t)ξ1 (t) − S12 P22 (t)ξ2 (t) dt ⎪ ⎪ ⎪ ⎪ ⎨   + Ap1 ξ1 (t) + Ap2 ξ2 (t) dw(t), (23a) ⎪ ⎪ ⎪ ⎪ ⎩ T P11 (t)ξ1 (t) − S22 P21 (t)ξ1 (t) − S22 P22 (t)ξ2 (t). (23b) 0 = A21 ξ1 (t) + A22 ξ2 (t) − S12 From (23b), we have ξ2 (t) = L(t)ξ1 (t).

(24)

Using the relation (24), after some calculations, we obtain    u∗ (t) = −R−1 B1T P11 (t) + B2T P21 (t) + P22 (t)L(t) ξ1 (t) −R

−1



B2T F (t)

Equation (25) can be rewritten as ⎡   ⎢ u∗ (t) = −R−1 B1T B2T ⎢ ⎣

(25)

 ξ2 (t) − L(t)ξ1 (t) . ⎤⎡ P11 (t)

P21 (t) + P22 (t)L(t) − F (t)L(t) 12

0



⎥ ⎢ ξ1 (t) ⎥ ⎥⎢ ⎥. ⎦⎣ ⎦ F (t) ξ2 (t)

(26)

Making an inverse transformation to (26) yields (19). By the same reason in the implementation of the optimal control as [6], we conclude that the two trajectories obtained by applying (22) and (26) respectively coincide exactly since A22 − S22 F (t) is invertible within the interval [0, T ]. Therefore, (19) constitutes the optimal linear feedback strategies for the problem (1)-(2) and their gains are not unique. Remark 2.5. Compared with [33], although the square completion technique adopted to prove the sufficient condition of well-posedness is the same, we chose a different Lyapunov function and obtained a new type of GDREs. Furthermore, we presented the existence condition of the solution of the GDREs. So, our results improved the known results in [33]. Remark 2.6. Unlike the standard stochastic systems, the optimal control strategy of stochastic singular systems is not unique.

3. LQ Pareto Optimal Control of Stochastic Singular Systems In this section, we consider the cooperative differential game that two players decide to coordinate their actions with an intent to minimize their cost functionals. For player i, i = 1, 2, the cost functional



 xT (t)Qi x(t) + uT1 (t)Ri1 u1 (t) + uT2 (t)Ri2 u2 (t) dt 0 + xT (T )E T QiT Ex(T ) ,

Ji (0, T, u1 , u2 , x0 ) = E

T

(27)

T T > 0, Rij = Rij ≥ 0, for j = i, j = 1, 2 and x(t) ∈ Rn is the where Qi = QTi , QiT = QTiT , Rii = Rii

state vector of the following stochastic singular equation ⎧   ⎪ ⎪ ⎨ Edx(t) = Ax(t) + B1 u1 (t) + B2 u2 (t) dt + Ap x(t)dw(t), ⎪ ⎪ ⎩ Ex(0) = x0 ,

(28)

where ui (t) ∈ Rmi is the control vector of player i. E, A, B1 , B2 , Ap are known matrices of appropriate dimensions and rank(E) = r ≤ n; x0 ∈ Rn is the compatible initial condition which is deterministic. Since the players coordinate their actions, we denote the joint action by u(t)  (u1 (t), u2 (t)) ∈ Rm with m =

2

i=1

mi . we assume an open-loop information structure.

The set of all admissible controls is denoted by U. In this section, we consider U = {u(·)  (u1 (·), u2 (·)) ∈ L2 [0, T ]| the corresponding solution x(·) of (28) satisfies x(·) ∈ L2 [0, T ]}. Since 13

we are interested in the joint minimization of the objectives of the players, the cost incurred by a single player cannot be minimized without increasing the cost incurred by other players. So, we consider solutions which cannot be improved upon by all the players simultaneously, i.e., the so-called Pareto optimal solutions. Definition 3.1. [9] Let U denote the set of admissible controls. Then uˆ ∈ U is called Pareto efficient if the set of inequalities ˆ, x0 ), i = 1, 2, Ji (0, T, u, x0 ) ≤ Ji (0, T, u does not allow for any solution u ∈ U, where at least one of the inequalities is strict. The corresponding point (J1 (0, T, u ˆ, x0 ), J2 (0, T, u ˆ, x0 )) ∈ R2 is called a pareto solution. The set of all Pareto solutions is called the Pareto frontier. The objective of this section is to find the set of Pareto efficient solutions of the finite horizon LQ differential game of the stochastic singular systems. First, we give the following assumption and lemmas.





⎢ E Assumption 3.1. rank ⎢ ⎣ 0

A

B1

E

0

B2 ⎥ ⎥ = n + rankE. ⎦ 0

Remark 3.1. Assumption 3.1 is in accordance with Assumption 2.2 when B = [B1 B2 ]. Likewise, Assumption 2.1 and Assumption 3.1 can guarantee the existence and uniqueness of the impulse-free solution of system (28). Under transformation (3), we can denote ⎡ ⎤ ⎢ B11 ⎥ ⎥, M B1 = ⎢ ⎣ ⎦ B12





⎢ B21 ⎥ ⎥, M B2 = ⎢ ⎣ ⎦ B22

then Assumption 3.1 is equivalent to that the matrix [A22

B12

(29)

B22 ] has full row rank, which

means that the system (28) is impulsive controllable. A well known way to find Pareto optimal controls is to solve a parameterized optimal control problem [16, 32]. Lemma 3.1, given below, states that every control minimizing a weighting sum of the cost functionals of all players (where all weights are in the unit simplex) is Pareto optimal. So, varying the weights over the unit simplex, one obtains, in principle, different Pareto optimal controls. A proof of the lemma can be found in [16].

14

Lemma 3.1. [16] Let α = (α1 , α2 ) ∈ A. Assume u ˆ ∈ U is such that u ˆ ∈ arg min

2 

u∈U

αi Ji (0, T, u, x0 ) ,

(30)

i=1

then u ˆ is Pareto efficient. So, we consider the weighting sum optimal control problem. For an arbitrary α = (α1 , α2 ) ∈ A, we define the weighting sum objective cost functional as Jα (0, T, u, x0 ) = E



T

0



 ¯ T Ex(T ) , ¯ ¯ xT (t)Qx(t) + uT (t)Ru(t) dt + xT (T )E T Q

(31)

¯ = 2 αi Qi , R ¯ = diag{α1 R11 + α2 R21 , α1 R12 + α2 R22 }, Q ¯T = where u(t) = [uT1 (t) uT2 (t)]T , Q i=1 2

i=1

αi QiT .

Accordingly, system (28) can be rewritten as ⎧   ⎪ ⎪ ¯ ⎨ Edx(t) = Ax(t) + Bu(t) dt + Ap x(t)dw(t),

(32)

⎪ ⎪ ⎩ Ex(0) = x0 ¯ = [B1 B2 ]. By (29), we have with B ¯= MB



 M B1

M B2



⎡ ⎢ B11 =⎢ ⎣ B12





¯1 ⎥ B21 ⎥ ⎢ B ⎥⎢ ⎥. ⎦ ⎣ ⎦ ¯2 B B22

(33)

Being a sufficient condition, it is unclear whether we obtain all Pareto optimal controls in this way. In fact, the above procedure may yield no Pareto optimal controls, while an infinite number of them can exist. The next example illustrates this point. Example 3.1. Consider the following stochastic singular system ⎧     ⎪ ⎪ ⎪ ⎪ dx1 (t) = x1 (t) + 2x3 (t) + u1 (t) dt + 3x2 (t) + x3 (t) dw(t), x1 (0) = 1, ⎪ ⎪ ⎨   dx2 (t) = u1 (t) − u2 (t) dt, x2 (0) = 0, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0 = x2 (t) + x3 (t)

(34)

with the cost functionals

J1 (0, 1, u1 , u2 ) =

0

J2 (0, 1, u1 , u2 ) =

1

1

0

[u1 (t) − u2 (t)]dt, x2 (t)[u2 (t) − u1 (t)]dt.

Then, by calculations, J2 (0, 1, u1 , u2 ) = − 12 x22 (1) = − 12 J12 (0, 1, u1 , u2 ) for all (u1 , u2 ). Obviously, by choosing different values of (u1 , u2 ), every point in curve J2 = − 21 J12 can be attained. Furthermore, by Definition 3.1, it is clear that every point on this curve is Pareto optimal. 15

Next, for 0 < α < 1, we consider the minimization of J(α)  αJ1 (0, 1, u1 , u2 ) + (1 − α)J2 (0, 1, u1 , u2 ) subject to (34). If we choose u1 (t) = c and u2 (t) = 0, straightforward calculations 2 yield that J(α) = αc − 1−α 2 c . By choosing c arbitrarily large, J(α) can be made arbitrarily small,

which means J(α) has not a minimum. Lemma 3.2 mentioned below points out that if the action spaces as well as the players’ objective functionals are convex, then minimization of the weighting sum of the objectives results in all Pareto solutions. Lemma 3.2. [10] Assume that the control space U is convex and the costs Ji (0, T, u, x0 ), i = 1, 2 are convex in u. Then, if u ˆ is Pareto efficient, there exist α ∈ A, such that u ˆ ∈ arg min u∈U

2 

αi Ji (0, T, u, x0 ) .

i=1

This property can be used to obtain both necessary and sufficient conditions for the existence of the Pareto optimal solutions. That is to say, under convexity conditions on Ji s and convexity assumption regarding the strategy space, all Pareto efficient strategies can be obtained by considering the minimization problem (30). To this end, we first derive a preliminary result. Lemma 3.3. Under Assumption 2.1 and Assumption 2.2, considering cost functional (2), where Q = QT , QT = QTT , R = RT and x(t) solves (1), J(0, T, u, x0 ) is convex as a function of u ∈ U, where U is any convex subset of Rm , if and only if J(0, T, u, 0) ≥ 0 for all u ∈ U. Proof. Let xu (t) denote the state trajectory of (1) driving by u. Assume v, w ∈ U. By convexity of strategy set U, for all λ ∈ [0, 1], we have λv + (1 − λ)w ∈ U. Under Assumption 2.1 and Assumption 2.2, the solution of the system (1) is unique. According to linearity property and uniqueness of the solution, we have xλv+(1−λ)w (t) = λxv (t) + (1 − λ)xw (t).

16

Using this relation, we obtain J(0, T, λv + (1 − λ)w, x0 ) T   xTλv+(1−λ)w Qxλv+(1−λ)w + (λv + (1 − λ)w)T R(λv + (1 − λ)w) dt =E 0 + xTλv+(1−λ)w (T )E T QT Exλv+(1−λ)w (T ) = λ2 J(0, T, v, x0 ) + (1 − λ)2 J(0, T, w, x0 ) T   + 2λ(1 − λ)E xTv Qxw + v T Rw dt + xTv (T )E T QT Exw (T ) . 0

Therefore, λJ(0, T, v, x0 ) + (1 − λ)J(0, T, w, x0 ) − J(0, T, λv + (1 − λ)w, x0 ) T   (xv − xw )T Q(xv − xw ) + (v − w)T R(v − w) dt = λ(1 − λ)E 0 T + (xv − xw ) (T )E T QT E(xv − xw )(T ) . Again, by the linearity property and the uniqueness of the solution, we identify the expression on the right side as λ(1 − λ)J(0, T, v − w, 0). The conclusion is proved. Combining Theorem 2.2 and Lemmas 3.1, 3.2 and 3.3, we can derive straightforwardly the existence and computational algorithms for the finite horizon Pareto optimal control problem. Lemma 3.4. Under Assumption 2.1 and Assumption 3.1, considering the cooperative game ¯ T and R = R, ¯ have a ¯ Q = Q, ¯ QT = Q (27) and (28). Assume that GDREs (9) with B = B, solution Pα (t) on [0, T ] for all α ∈ A. Then, all Pareto efficient solutions are obtained by u∗α := arg min(α1 J1 + α2 J2 ), subject to (32) u∈U

  via determining α ∈ A. The corresponding Pareto solutions are J1 (u∗α ), J2 (u∗α ) . Proof. Under the assumption that the GDREs (9) admit a solution Pα (t) on [0, T ] with ¯ T and R = R, ¯ according to Theorem 2.2, the finite horizon weighting sum ¯ Q = Q, ¯ QT = Q B = B, optimal control problem (31) and (32) is well-posed, and the cost functional (31) attains a minimum Jα (0, T, u∗ , x0 ) = xT (0)E T Pα (0)x(0). So, Jα (0, T, u∗, 0) = 0. By Lemma 3.3, under Assumption 2.1 and Assumption 3.1, Jα (0, T, u, x0 ) is convex in u for all α ∈ A. Therefore, by Lemmas 3.1 and 3.2, all Pareto efficient solutions can be obtained by the weighting sum optimization method. This yields the next procedure to calculate all Pareto efficient solutions for this game. Theorem 3.1. Under Assumption 2.1 and Assumption 3.1, considering the cooperative game

17

(27) and (28). Assume that (35) below admits a solution on [0, T ] for arbitrary α ∈ A. ⎧ ¯ + AT (E + )T P T (t)EE + Ap −E T P˙ (t) = AT P (t) + P T (t)A + Q ⎪ p ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ¯ T P (t), ¯R ¯ −1 B ⎪ − P T (t)B ⎨ ⎪ ⎪ ¯ T E, ⎪ E T P (T ) = E T Q ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ E T P (t) = P T (t)E.

(35b) (35c)

Then, the set of all cooperative Pareto solutions is given by family of the Pareto efficient strategies is ⎡ ⎢ u∗α (t) = ⎢ ⎣ where

u∗1α (t) u∗2α (t)

{(J1 (u∗α ), J2 (u∗α ))|α

∈ A}. Here, the

⎤ ⎥ ¯ T Kα (t)x(t), ¯ −1 B ⎥ = −R ⎦



⎡ ⎢ Kα (t) = M T ⎢ ⎣

(35a)

Pα11 (t)

0

Pα21 (t) + Pα22 (t)Lα (t) − F¯ (t)Lα (t) F¯ (t)



⎥ −1 ⎥N ⎦



⎢ Pα11 (t) with Pα (t) = M T ⎢ ⎣ Pα21 (t)

Pα12 (t) ⎥ ⎥ N −1 being a solution of (35). Denoting ⎦ Pα22 (t) ⎡

⎤ ⎡ ¯ ⎢ B1 ⎥ −1 ⎢ ¯ ⎢ ⎥R S¯ = ⎢ ⎣ ⎦ ⎣ ¯ B2

⎤T ⎡ ¯ B1 ⎥ ⎢ S¯11 ⎥ ⎢ ⎦ ⎣ ¯2 B S¯21

⎤ ¯ S12 ⎥ ⎥, ⎦ ¯ S22

T then Lα (t) = −[A22 − S¯22 Pα22 (t)]−1 [A21 − S¯12 Pα11 (t) − S¯22 Pα21 (t)] and F¯ (t) is any time-varying

¯ α (t), the matrix making A22 − S¯22 F¯ (t) invertible within the interval [0, T ]. With Acl (t) = A − SK closed-loop system is Edx(t) = Acl (t)x(t)dt + Ap x(t)dω(t), Ex(0) = x0 . The corresponding cost is Ji (u∗α ) = xT (0)E T P˜iα (0)x(0), where P˜iα (t) is a solution of the following GDREs ⎧ ⎪ ⎪ T T ⎪ −E T P˜˙iα (t) = ATcl (t)P˜iα (t) + P˜iα (t)Acl (t) + Qi + ATp (E + )T P˜iα (t)EE + Ap ⎪ ⎪ ⎪ ⎡ ⎤ ⎪ ⎪ ⎪ ⎪ ⎪ 0 ⎥ ⎪ ⎢ Ri1 ⎪ ⎪ ⎥R ¯ −1 B ¯ T Kα (t), ¯R ¯ −1 ⎢ + KαT (t)B ⎨ ⎣ ⎦ 0 Ri2 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ E T P˜iα (T ) = E T QiT E, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ T ⎩ E T P˜iα (t) = P˜iα (t)E. The proof is similar to Theorem 2.2, so we omit it due to the limitation of space. Remark 3.2. Although the basic idea (under convexity conditions on cost functionals and convexity assumption regarding the strategy space, all Pareto efficient strategies can be obtained 18

by considering the weighting sum minimization problem) is the same with [9], our system is more complicated due to the presence of singular matrix E and Brownian motion w(t). So, our conclusion generalized the known result in [9], which mainly studied the Pareto optimal control of deterministic systems. Besides, the theory of the related LQ optimal control of stochastic singular systems is immature, which increased the research challenge.

4. Examples In order to demonstrate the efficiency of the proposed approach, we consider the following ecological fishing model (see [11], Section 4), which is described by the following dynamic equations: f˙(t)

=

βf (t) − p(t) − u1 (t), f (0) = f0 ,

(36)

εp(t) ˙

=

γp(t) + u2 (t),

(37)

where f is the fish stock, p is the amount of pollution, u1 is the fishing by the ranger, u2 is the waste dropped into the river, β and γ are nonzero constant coefficients. The revenue functionals to be maximized are

J1

= =

  τf f 2 (t) − u21 (t) dt,

T

  2 τp p (t) − u22 (t) dt.

0

J2

T

0

Letting ε = 0 in (37), we get an algebraic equation, it means that the negative effect of pollution on the growth of the fish stock is proportional to p. Introducing xT (t) := [f T (t) pT (t)]T , we can rewrite the above model (36)-(37) as Ex(t) = Ax(t) + B1 u1 (t) + B2 u2 (t), which is a deterministic singular system, where ⎡ ⎤ ⎡ ⎤ ⎢ 1 0 ⎥ ⎥, E=⎢ ⎣ ⎦ 0 0

⎢ β A=⎢ ⎣ 0



−1 ⎥ ⎥, ⎦ γ



⎢ −1 ⎥ ⎥, B1 = ⎢ ⎣ ⎦ 0





⎢ 0 ⎥ ⎥. B2 = ⎢ ⎣ ⎦ 1

However, the white noise always exists and can not be omitted. As a result, we introduce stochastic perturbation to the fish stock f and obtain the following stochastic singular system:   Edx(t) = Ax(t) + B1 u1 (t) + B2 u2 (t) dt + Ap x(t)dw(t), 19





⎢ σ where Ap = ⎢ ⎣ 0

0 ⎥ ⎥ with σ representing the intensity of the white noise, which is in the form ⎦ 0

of system (28). Obviously, Assumption 2.1 and Assumption 3.1 are satisfied in this example. Accordingly, the revenue functions can be rewritten as the form in (27) with ⎡ ⎡ ⎡ ⎤ ⎤ ⎤ ⎢ −τf Q1 = ⎢ ⎣ 0

0 ⎥ ⎥, ⎦ 0

⎢ 0 Q2 = ⎢ ⎣ 0

0 −τp

R11 = R22 = 1, So



⎢ −α1 τf ⎢ −1 0 ⎥ ⎥, Q ¯=⎢ ¯ =⎢ B ⎣ ⎦ ⎣ 0 1 0

⎢ 0 Q1T = Q2T = ⎢ ⎣ 0

0 ⎥ ⎥, ⎦ 0

R12 = R21 = 0.







⎥ ⎥, ⎦



⎢ α1 ⎥ ⎥, R ¯=⎢ ⎣ ⎦ −α2 τp 0 0







0 ⎥ ⎢ 0 ⎥, Q ¯T = ⎢ ⎣ ⎦ 0 α2

0 ⎥ ⎥ ⎦ 0

in the cooperative differential game (31)-(32). In what follows, we will verify the effectiveness of Theorem 3.1. For any α = (α1 , α2 ) ∈ A, although Assumption 2.3 is not satisfied, when γ 2 > τp and b2 − 4ac > 0, where a =

1 α1

+

1 α2

1 γ 2 −τp ,

·

b = σ 2 + 2β, c = α1 τf , there exists a solution to

the GDREs (35). By straightforward calculations, we obtain the following solution of the involved GDREs (35) ⎤

⎡ ⎢ Pα11 Pα = ⎢ ⎣ Pα21 where d1 = b +



Pα12 ⎥ ⎢ ⎥=⎢ ⎦ ⎣ −√ Pα22

d1 a(t−T ) ] 2a [1+e d1 a(t−T ) 1+ d e 2

1 γ 2 −τp

·

d1 a(t−T ) ] 2a [1+e d 1+ d1 ea(t−T ) 2

⎤ 0 α2 γ + α2

 γ 2 − τp

⎥ ⎥, ⎦

√ √ b2 − 4ac and d2 = b − b2 − 4ac. Elementary manipulations show then that the

family of the Pareto efficient strategies is u∗1α

=

u∗2α

=

1 Pα11 f ∗ , α1 1 − [(Pα21 + Pα22 Lα − F¯ Lα )f ∗ + F¯ p∗ ], α2

where Lα = (α2 γ−Pα22 )−1 Pα21 and F¯ = α2 γ is a constant, f ∗ solves df ∗ = (β−Lα − α11 Pα11 )f ∗ dt+ σf ∗ dw, f ∗ (0) = f0 and p∗ = Lα f ∗ . When F¯ = 0, we plotted the Pareto frontier in Fig.1 for the parameter values β = 3, γ = 2, σ = 0.2, τf = 1, τp = 1, T = 5 and f0 = 1. If the weight of player 1 in the cooperative cost functional would be α1 = 0.6, the state and control trajectories are shown in Fig.2 and Fig.3, respectively. 20

1.4

1.2

1

* α

J2(u )

0.8

0.6

0.4

0.2

0 −0.15

−0.1

−0.05

0

J1(u* ) α

Fig.1. Pareto frontier when F¯ = 0.

1.6 f* *

p

1.4

1

*

Optimal states f and p

*

1.2

0.8 0.6 0.4 0.2 0

0

0.5

1

1.5 Time(sec)

2

Fig.2. State trajectories if α1 = 0.6, when F¯ = 0.

21

2.5

3

5 *

u1α 4

*

u2α

*

*

Optimal controls u1α and u2α

3 2 1 0 −1 −2 −3 −4

0

0.5

1

1.5 Time(sec)

2

2.5

3

Fig.3. Control trajectories if α1 = 0.6, when F¯ = 0.

5. Conclusions In this paper, we have investigated the finite horizon LQ Pareto optimal control of stochastic singular systems. Firstly, finite horizon stochastic singular LQ optimal control has been discussed by means of the square completion technique, and a new type of GDREs have also been established. The existence condition of the solution of the GDREs has been put forward. Secondly, for finite horizon LQ Pareto optimal control, it is shown that the solvability of the corresponding GDREs presents a sufficient condition, under which all Pareto candidates can be obtained by solving a weighting sum optimal control. A numerical example has also been provided to show the effectiveness of our main results. The event-triggering mechanism has the advantage of saving the communication and computation resources. In our future research, the event-triggering mechanism [17, 18] will be taken into account in the analysis of optimal control of stochastic singular systems.

References [1] D. Acemoglu, Introduction to Modern Economic Growth, Princeton University Press, 2008. [2] M. Ait Rami, J.B. Moore, X.Y. Zhou, Indefinite stochastic linear quadratic control and generalized differential Riccati equation, SIAM Journal on Control and Optimization 40 (2001) 22

1296-1311. [3] T. Basar, G.J. Olsder, Dynamic Noncooperative Game Theory, SIAM Series in Classics in Applied Mathematics, 1999. [4] A. Bensoussan, Lectures on stochastic control-Part I, in: Lecture Notes in Mathematics: Springer-Verlag, 972 (1983) 1-39. [5] E.K. Boukas, Stabilization of stochastic singular nonlinear hybrid systems, Nonlinear Analysis 64 (2006) 217-228. [6] D. Cobb, Descriptor variable systems and optimal state regulation, IEEE Transactions on Automatic Control 28(5) (1983) 601-611. [7] L. Dai, Singular Control Systems, Lecture Notes in Control and Information Sciences, Springer-Verlag, New York, 1989. [8] E.J. Dockner, S. Jorgensen, N.V. Long, G. Sorger, Differential Games in Economics and Management Science, Cambridge University Press, 2001. [9] J.C. Engwerda, The regular convex cooperative linear quadratic control problem, Automatica 44(9) (2008) 2453-2457. [10] J.C. Engwerda, Necessary and sufficient conditions for Pareto optimal solutions of cooperative differential games, SIAM Journal on Control and Optimization 48(6) (2010) 3859-3881. [11] J.C. Engwerda, Salmah, Feedback Nash equilibria for linear quadratic descriptor differential games, Automatica 48(4) (2010) 625-631. [12] F.R. Gantmacher, The Theory of Matrices, AMS Chelsea Publishing, 1959. [13] Z.W. Gao, X.Y. Shi, Observer-based controller design for stochastic descriptor systems with Brownian motions, Automatica 49(7) (2013) 2229-2235. [14] M. Gerdin, J. Sjoberg, Nonlinear stochastic differential-algebraic equations with application to particle filtering, 2006 45th IEEE Conference on Decision and Control (2006) 6630-6635. [15] S.G. Hu, B.M. Huang, F.K. Wu, Stochastic differential equations, Science Press, Beijing, 2008. 23

[16] G. Leitmann, Cooperative and Noncooperative Many Player Differential Games, Springer Verlag, Berlin, 1974. [17] Q. Li, B. Shen, J.L. Liang, H.S. Shu, Event-triggered synchronization control for complex networks with uncertain inner coupling, International Journal of General Systems 44(2) (2015) 212-225. [18] Q. Li, B. Shen, Y.R. Liu, F.E. Alsaadi, Event-triggered H∞ state estimation for discrete-time stochastic genetic regulatory networks with Markovian jumping parameters and time-varying delays, Neurocomputing 174 (2016) 912-920. [19] S.P. Ma, C.H. Zhang, Z.L. Cheng, Delay-dependent robust H∞ control for uncertain discretetime singular systems with time-delay, Journal of Computational and Applied Mathematics 217(1) (2008) 194-211. [20] X.R. Mao, C.G. Yuan, Stochastic Differential Equations with Markovian Switching, Imperial College Press, London, 2006. [21] X.R. Mao, Stochastic Differential Equations and Applications, Horwood Publishing, Chichester, 2007. [22] F.P. Ramsey, A mathematical theory of saving, The Economic Journal 38(152) (1928) 543-559. [23] P.V. Reddy, J.C. Engwerda, Pareto optimality in infinite horizon linear quadratic differential games, Automatica 49(6) (2013) 1705-1714. [24] P.V. Reddy, J.C. Engwerda, Necessary and sufficient conditions for Pareto optimality in infinite horizon cooperative differential games, IEEE Transactions on Automatic Control 59(9) (2014) 2536-2542. [25] O. Schein, G. Denk, Numerical solution of stochastic differential-algebraic equations with applications to transient noise simulation of microelectronic circuits, Journal of Computational and Applied Mathematics 100(1) (1998) 77-92. [26] P. Shi, E.K. Boukas, On H∞ control design for singular continuous-time delay systems with parametric uncertainties, Nonlinear Dynamics and Systems Theory 4(1) (2004) 59-71. 24

[27] G. Wang, Stochastic stabilization of singular systems with Markovian switchings, Applied Mathematics and Computation 250 (2015) 390-401. [28] Y.Y. Wang, S.J. Shi, Z.J. Zhang, A descriptor-system approach to singular perturbation of linear regulators, IEEE Transactions on Automatic Control 33(4) (1988) 370-373. [29] Y.Q. Xia, E.K. Boukas, P. Shi, J.H. Zhang, Stability and stabilization of continuous-time singular hybrid systems, Automatica 45(6) (2009) 1504-1509. [30] H. Xu, K. Mizukami, On linear-quadratic optimal regulator for continuous-time descriptor system, Proceedings of the 31st IEEE Conference on Decision and Control (1992) 987-988. [31] S. Xu, J. Lam, Robust stability and stabilization of discrete singular systems: An equivalent characterization, IEEE Transactions on Automatic Control 49(4) (2004) 568-574. [32] L.A. Zadeh, Optimality and non-scalar-valued performance criteria, IEEE Transactions on Automatic Control 8(1) (1963) 59-60. [33] Q.L. Zhang, S.Y. Xing, L. Sheng, Stability analysis and optimal control of stochastic singular systems, Optimization Letters 8(6) (2014) 1905-1920. [34] W.H. Zhang, B.S. Chen, On stabilizability and exact observability of stochastic systems with their applications, Automatica 40(1) (2004) 87-94. [35] W.H. Zhang, Y. Zhao, L. Sheng, Some remarks on stability of stochastic singular systems with state-dependent noise, Automatica 51 (2015) 273-277. [36] F. Zhao, Q.L. Zhang, Y. Zhang, H∞ filtering for a class of singular biological systems, IET Control Theory and Applications 9 (2015) 2047-2055. [37] J.D. Zhu, S.P. Ma, Z.L. Cheng, Singular LQ problem for nonregular descriptor systems, IEEE Transactions on Automatic Control 47(7) (2002) 1128-1133. [38] S.Q. Zhu, C.H. Zhang, Z.L. Cheng, J.E. Feng, Delay-dependent robust stability criteria for two classes of uncertain singular time-delay systems with multiple delays, IEEE Transactions on Automatic Control 52(5) (2007) 1188-1191.

25