Abstract: In this paper, a framework is established to model a deceitful agent and train it in an adversarial two-player game. In the game, a player uses multi-objective deception to manipulate its opponent's belief about its true intention. In this regard, the player is trained to switch between different strategies based on its state in the game. There is a lower-level policy, which stores the policy to carry out a primitive task. Moreover, there is a higher-level policy, which changes the desired task in different states. The game of guarding territories is utilized to investigate the control mechanism. The lower-level policy is trained via the fuzzy actor-critic learning (FACL) algorithm, and the higher-level policy is extracted via the non- dominated sorting genetic algorithm II (NSGA-II). The results show that by implementing a two-level policy, the invader can increase its pay-off against a non-deceptive situation. In addition, a comparison is conducted between the higher-level policies based on their input information. PDF
Keywords: Differential Games, Reinforcement Learning, Actor-Critic Learning, Fuzzy Systems