In this article, we show how to obtain all of the Pareto optimal decision vectors and solutions for the finite horizon indefinite mean-field stochastic cooperative linear-quadratic (LQ) difference game. First, the equivalence between the solvability of the introduced $\mathcal {N}$ coupled generalized difference Riccati equations (GDREs) and the solvability of the multiobjective optimization problem is established. However, it is difficult to obtain Pareto optimal decision vectors based on the $\mathcal {N}$ coupled GDREs because the optimal joint strategy adopted by all players to optimize the performance criterion of some players in the game is different from the strategies of other players, which rely on the weighted matrices of cost functionals that may be different among players. Second, a necessary and sufficient condition is developed to guarantee the convexity of the costs, which makes the weighting technique not only sufficient but also necessary for searching Pareto optimal decision vectors. It is then shown that the mean-field Pareto optimality algorithm (MF-POA) is presented to identify, in principle, all of the Pareto optimal decision vectors and solutions via the solutions to the weighted coupled GDREs and the weighted coupled generalized difference Lyapunov equations (GDLEs), respectively. Finally, a cooperative network security game is reported to illustrate the results presented. Simulation results validate the solvability, correctness, and efficiency of the proposed algorithm.
23 January 2020 in IEEE Transactions on Systems, Man, and Cybernetics: Systems
This article derives necessary and sufficient conditions for Pareto optimal solutions in infinite horizon cooperative difference games of autonomous systems with exponentially discounted performances. First, the $\mathcal {N}$ constrained optimal control problems are converted into an unconstrained characterization with mixed endpoint constraints by introducing appropriate auxiliary states. Second, the infinite horizon cooperative difference games are transformed equivalently into an augmented and truncated finite horizon optimization problem by defining two real-valued functions and the necessary conditions are derived based on the maximum principle (MP) of discrete-time type. Moreover, we present sufficient conditions for general nonautonomous systems. Finally, the results developed are employed to address the linear quadratic (LQ) cooperative difference games for fixed as well as arbitrary initial states.
30 August 2022 in IEEE Transactions on Automatic Control
This paper is concerned with the mean-field stochastic cooperative linear-quadratic (LQ) dynamic difference game in an infinite time horizon. First, the necessary and sufficient conditions for the stability in the mean-square sense, and the stochastic Popov-Belevith-Hautus (PBH) eigenvector tests for exact observability and exact detectability of mean-field stochastic linear difference systems are derived by the $\mathscr {H}$-representation technique. Second, the relation between the solvability of the cross-coupled generalized Lyapunov equations (CC-GLEs) and exact observability, exact detectability, and stability of the mean-field dynamic system is well characterized. It is then shown that the cross-coupled algebraic Riccati equations (CC-AREs) admit a unique positive definite (positive semi-definite, respectively) solution under exact observability (exact detectability, respectively), which is also a feedback stabilizing solution. Furthermore, all Pareto optimal strategies and solutions can be respectively derived via the solutions to the weighted CC-AREs (WCC-AREs) and the weighted cross-coupled algebraic Lyapunov equations (WCC-ALEs). Finally, a practical application on the computation offloading in the multi-access edge computing network (MECN) is presented to illustrate the proposed theoretical results.
1 October 2021 in Journal of the Franklin Institute