We compare the performance of Inverse Reinforcement Learning (IRL) with therelative new model of Multi-agent Inverse Reinforcement Learning (MIRL). Beforecomparing the methods, we extend a published Bayesian IRL approach that is onlyapplicable to the case where the reward is only state dependent to a generalone capable of tackling the case where the reward depends on both state andaction. Comparison between IRL and MIRL is made in the context of an abstractsoccer game, using both a game model in which the reward depends only on stateand one in which it depends on both state and action. Results suggest that theIRL approach performs much worse than the MIRL approach. We speculate that theunderperformance of IRL is because it fails to capture equilibrium informationin the manner possible in MIRL.
展开▼
