Acquisition of cooperative behaviour among heterogeneous agents using step-up reinforcement learning

This paper discusses acquisition of cooperative behaviour among heterogeneous agents and proposes two methods to promote cooperative behaviour: phased learning and selective recognition. For complicated scenarios such as multi-agent tasks, we propose phased learning, in which agents first learn in a simpler environment before learning in the target environment. For heterogeneous multi-agent tasks, we propose selective recognition, in which an agent recognizes a partner, with whom it can cooperate to earn rewards, selectively. By means of simulations in which two types of agents cooperated to capture prey, we verified that, using our proposed methods, agents are able to differentiate agents they should cooperate with from those with whom they should not.