Search

Each agent takes the others different from the environment, and maintains the others' replacing eligibility traces to estimate the other's behaviors during learning, at the same time estimates itself behavior. Each agent does not know the others' structures of Q function and payoff using this multi agent reinforcement learning without rigorous condition. The simulation results prove the validity of this proposed learning method and the high efficiency according to the centralized Q-learning. [ABSTRACT FROM AUTHOR]

By adopting control method based on behavior, mobile robot can accomplish its given task using reinforcement learning while need not know the exact environment mode. The robot needs to memorize the behaviors and states, at the same time the EMS memory space is not enough, the connectionist reinforcement learning can approximate the Q function with MLPs, and generalize the state space to economize the memory space. The results of simulation show that the algorithm is valid, and it resolves the shortcoming that the reinforcement learning with look up table is not fit for the continuous state space. This method provides the bases for the reality of mobile robot. [ABSTRACT FROM AUTHOR]