Extended stochastic reinforcement learning for the acquisition of cooperative motion plans for dynamically constrained agents

This paper examines the problem of the acquisition of coordinated task plans for a group of autonomous agents. The authors deal with cases when intelligent mobile robots are on a seesaw and they are trying to balance the seesaw. The objective of the agents is to maximize the global optimization function under the constraints that the effect of their decision is propagated after a certain time delay. To cope with such a situation, this paper proposes extensions to the learning automata type reinforcement learning methods. One is the group learning method. It generates teaching signals that are robust to the delay of the result of an action. Another is the genetic reinforcement learning phase. This is intended to give hard-wired knowledge to all the agents through the meta-learning phase. A sensor information compression function is acquired as the knowledge and a genetic algorithm is used for the search mechanism. The authors demonstrate how the cooperative plans can be acquired for seesaw balancing problem where conventional reinforcement learning could not achieve its balance