Multi-agent Deception in Attack-Defense Stochastic Game

Abstract

AbstractThis paper studies a sequential adversarial incomplete information game, the attack-defense game, with multiple defenders against one attacker. The attacker has limited information on game configurations and makes guesses of the correct configuration based on observations of defenders’ actions. Challenges for multi-agent incomplete information games include scalability in terms of agents’ joint state and action space, and high dimensionality due to sequential actions. We tackle this problem by introducing deceptive actions for the defenders to mislead the attacker’s belief of correct game configuration. We propose a k-step deception strategy for the defender team that forward simulates the attacker and defenders’ actions within k steps and computes the locally optimal action. We present results based on comparisons of different parameters in our deceptive strategy. Experiments show that our approach outperforms Bayesian Nash Equilibrium strategy, a strategy commonly used for adversarial incomplete information games, with higher expected rewards and less computation time.KeywordsAttack-defense gameMultiple agent systemsDefense strategyGame theoryZero-sum gameGames of incomplete informationDeception