A multiagent reinforcement learning method is presented in this paper, along with its application to the domain of robot soccer. Reinforcement learning (RL) in a multiagent system (MAS) should be redefined to consider the joint-state and joint-action of all agents, while taking into account that other agents' actions can only be predicted. The authors propose using a probabilistic neural network (PNN) to predict actions based on joint-states as inputs. Moreover, they propose the use of a fuzzy Q-learning algorithm for RL to reduce the dimensionality of the mapping from the state space to the action space.
展开▼
