Abstract/Summary

Whereas most work in reinforcement learning (RL) ignores the structure or relationships between actions, in this paper we show that exploiting structure in the action space can improve sample efficiency during exploration. To show this we focus on concurrent action spaces where the RL agent selects multiple actions per timestep. Concurrent action spaces are challenging to learn in especially if the number of actions is large as this can lead to a combinatorial explosion of the action space. This paper proposes two methods: a first approach uses implicit structure to perform high-level action elimination using task-invariant actions; a second approach looks for more explicit structure in the form of action clusters. Both methods are context-free, focusing only on an analysis of the action space and show a significant improvement in policy convergence times.