アイテム詳細

要旨

We investigate how reinforcement learning agents can learn tocooperate. Drawing inspiration from human societies, in whichsuccessful coordination of many individuals is often facilitated byhierarchical organisation, we introduce Feudal Multi-agent Hierar-chies (FMH). In this framework, a ‘manager’ agent, which is taskedwith maximising the environmentally-determined reward func-tion, learns to communicate subgoals to multiple, simultaneously-operating, ‘worker’ agents. Workers, which are rewarded for achiev-ing managerial subgoals, take concurrent actions in the world. Weoutline the structure of FMH and demonstrate its potential for de-centralised learning and control. We find that, given an adequate setof subgoals from which to choose, FMH performs, and particularlyscales, substantially better than cooperative approaches that use ashared reward function.