Exploration via State influence Modeling

This paper studies the challenging problem of reinforcement learning (RL) in hard exploration tasks with sparse rewards. It focuses on stage before agent gets first positive reward, which case, traditional RL algorithms simple strategies often work poorly. Unlike previous methods using some attribute a single state as intrinsic reward to encourage exploration, this leverages social influence between different states permit more efficient exploration. introduces general construction method evaluate dynamically. Three kinds are introduced for state: conformity, power, and authority. By measuring state’s influence, agents quickly find focus during process. The proposed framework evaluation works well task. Extensive experimental analyses comparisons Grid Maze many Atari 2600 games demonstrate its high efficiency.