An Object Oriented Approach to Fuzzy Actor-Critic Learning for Multi-Agent Differential Games

摘要

This paper presents a new form of the multi-agent fuzzy actor-critic learning algorithm for differential games. An object oriented approach to defining the relationships between agents is proposed. We define the fuzzy inference system as a network structure and define attributes of the agents as rule sets that fired and rewards associated with the fired rule set. The resulting fuzzy actor-critic reinforcement learning algorithm is investigated for playing the differential pursuer super evader game. The game is played in a continuous state and action space to simulate a real world environment. All the robots in the game are simultaneously learning.