Cooperative multi-agent reinforcement learning in a large stationary environment

摘要

Reinforcement learning comprises an attractive solution to the multi-agent cooperation problem, due to its robustness for learning in unknown and uncertain environments. The objective of this paper is to provide learning capabilities to a group of autonomous agents in order to efficiently perform a cooperative foraging task in a distributed manner. Firstly, the D-DCM-MultiQ learning method, presented in [1], is evaluated. To overcome the shortcomings of this method, new cooperative action selection strategies are developed. A new exploration alternative, favoring least recently visited states, is also proposed. The conducted simulation tests indicate the efficiency of suggested improvements in the case of large, unknown and stationary environments.