The Stackelberg equilibrium for one-sided zero-sum partially observable stochastic games. (June 2022)

• Record Type: Journal Article
• Authors: Zheng, Wei; Jung, Taeho; Lin, Hai
• Is Part Of: Automatica. Volume 140(2022)
• Publication Date: 2022-06
• Subjects: Game theories -- Stochastic control and game theory -- Dynamic games -- Optimization under uncertainties; Automatic control -- Periodicals; Automation -- Periodicals; 629.805
• DOI: 10.1016/j.automatica.2022.110231
• Languages: English
• ISSNs: 0005-1098

Abstract

Abstract: Formulating cyber-security problems with attackers and defenders as a partially observable stochastic game has become a trend recently. Among them, the one-sided two-player zero-sum partially observable stochastic game (OTZ-POSG) has emerged as a popular model because it allows players to compete for multiple stages based on partial knowledge of the system. All existing work on OTZ-POSG has focused on the simultaneous move scenario and assumed that one player's actions are private in the execution process. However, this assumption may become questionable since one player's action may be detected by the opponent through deploying action detection strategies. Hence, in this paper, we propose a turn-based OTZ-POSG with the assumption of public actions and investigate the existence and properties of a Stackelberg equilibrium for this game. We first prove the existence of the Stackelberg equilibrium for the one-stage case and show that the one-stage game can be converted into a linear-fractional programming problem and therefore solved by linear programming. For multiple stages, the main challenge is the information leakage issue as the public run-time action reveals certain private information to the opponent and allows the opponent to achieve more rewards in the future. To deal with this issue, we adopt the concept of ε -Stackelberg equilibrium and prove that this equilibrium can be achieved for finite-horizon OTZ-POSGs. We propose a space partition approach to solve the game iteratively and show that the value function of the leader is piece-wise linear and the value function of the follower is piece-wise constant for multiple stages. Finally, examples are given to illustrate the space partition approach and show that value functions are piece-wise linear and piece-wise constant.