準エキスパート集団からのアンサンブル逆強化学習

抄録

Ensemble inverse reinforcement learning from semi-experts' behavior is proposed. In many inverse reinforcement learning (IRL) problem, the expert agent which has ideal rewards for achieving the goal is supposed to be existing. However, in real world problem, the expert is not always observed. Moreover, the estimated reward function includes the bias depending on its inherent behavior if the reward for achieving the goal task is estimated from one agent. In order to overcome the limitation of IRL, we apply Adaboost, one of ensemble and boosting approach, to IRL and integrate estimated reward functions from semi-expert agents. To confirm the effectiveness of the proposed method in the grid world including incomplete areas, we compared the results of reinforcement learning using estimated reward functions and integrated reward function by simulation. The simulation result shows the proposed method can estimate the reward adaptively.