An actor-critic approach for learning cooperative behaviors of multiagent seesaw balancing problems

This paper proposes a new approach to realize a reinforcement learning scheme for autonomous multiple agents system. In our approach, we treat the cooperative agents systems in which there are multiple autonomous mobile robots, and the seesaw balancing task is given. This problem is an example of corresponding tasks to find the appropriate locations for multiple mobile robots. Each robot agent on a seesaw keeps being balanced state. As a most useful algorithm, the Q-learning method is well known. However, feasible action values of robot agents must be categorized into some discrete action values. Therefore, in this study, the actor-critic method is applied to treat continuous values of agents´ actions. Each robot agent has a set of normal distribution, that determines a distance of the robot movement for a corresponding state of the seesaw system. Based on a result of movement in this system, the normal distribution is modified by actor-critic learning method. The simulation result shows the effectiveness of our approaching method.