Centralized Norm Enforcement in Mixed-Motive Multiagent Reinforcement Learning.

Mixed-motive games comprise a subset of games in which individual and collective incentives are not entirely aligned. These games are relevant because they frequently occur in real-world and artificial societies, and their outcome is often bad for the involved parties. Institutions and norms offer a good solution for governing mixed-motive systems. Still, they are usually incorporated into the system in a distributed fashion, or they are not able to dynamically adjust to the needs of the environment at run-time. We propose a way of reaching socially good outcomes in mixed-motive multiagent reinforcement learning settings by enhancing the environment with a normative system controlled by an external reinforcement learning agent. By adopting this proposal, we show it is possible to reach social welfare in a mixed-motive system of self-interested agents using only traditional reinforcement learning agent architectures.