Abstract
This paper proposes a design method for a multi-robot system that acquires cooperative behaviors autonomously. From embodied cognitive point of view, a robot should have the capability of acquiring appropriate cooperative behaviors through its experience. One of the most important issues for this function is how to design an on-line autonomous behavior acquisition mechanism capable of developing the robot's role in an embedded environment. In our approach, reinforcement learning that uses Bayesian discrimination method for segmenting the continuous state and action spaces simultaneously is applied to a robot for behavior acquisition. In addition to this, in order to support the learning in a dynamic environment that originates from the other learning robots, neural networks are provided for predicting the other robots' moves at the next time step. The output signals are utilized as the sensory information of the reinforcement learning to increase the stability of the learning problem. A homogeneous real robot system will be built for evaluating our approach.
