Learning to Collaborate from Delayed Rewards in Foraging Like Environments

Machine learning techniques are usually used in coordination problems and in competitive games but not in collaborative ones. Collaboration and coordination are different, while in coordination the task can not be concluded by a unique agent, in collaboration it can be solved by one agent or by a team, but the use of several agents has to be re ected in the performance of the system. In this work, authors propose the use of in uence value reinforcement learning (IVRL) in collaborative problems and test it into a foraging game. In early work authors show experimentally that, in coordination problems, the IVRL paradigm performs better than the traditional paradigms (independent learning and joint action learning). Thus, in this paper, authors compare their new paradigm (IVRL) with the traditional ones in order to establish if reinforcement learning is well suited to be used in collaboration problems, and shows that the proposed paradigm performs better than the traditional ones.