Fuzzy policy gradient reinforcement learning for leader-follower systems

This paper presents a policy gradient multi-agent reinforcement learning algorithm for leader-follower systems. In this algorithm, cooperative dynamics of the leader-follower control is modelled as an incentive Stackelberg game. A linear incentive mechanism is used to connect the leader and follower policies. Policy gradient reinforcement learning explicitly explores policy parameter space to search the optimal policy. Fuzzy logic controllers are used as the policy. The parameters of fuzzy logic controllers can be improved by this policy gradient algorithm.