Learning to Factor Policies and Action-Value Functions: Factored Action Space Representations for Deep Reinforcement learning
This addresses the challenge of inefficient learning in high-dimensional visual decision-making domains with structured action spaces, offering a novel paradigm for reinforcement learning agents.
The paper tackles the problem of exploiting compositional structure in discrete action spaces for deep reinforcement learning by proposing Factored Action space Representations (FAR), which decomposes control policies into independent components, resulting in improved performance over baseline methods in Atari 2600 tasks, with FARA3C outperforming A3C in 9 out of 14 tasks and FARAQL outperforming AQL in 9 out of 13 tasks.
Deep Reinforcement Learning (DRL) methods have performed well in an increasing numbering of high-dimensional visual decision making domains. Among all such visual decision making problems, those with discrete action spaces often tend to have underlying compositional structure in the said action space. Such action spaces often contain actions such as go left, go up as well as go diagonally up and left (which is a composition of the former two actions). The representations of control policies in such domains have traditionally been modeled without exploiting this inherent compositional structure in the action spaces. We propose a new learning paradigm, Factored Action space Representations (FAR) wherein we decompose a control policy learned using a Deep Reinforcement Learning Algorithm into independent components, analogous to decomposing a vector in terms of some orthogonal basis vectors. This architectural modification of the control policy representation allows the agent to learn about multiple actions simultaneously, while executing only one of them. We demonstrate that FAR yields considerable improvements on top of two DRL algorithms in Atari 2600: FARA3C outperforms A3C (Asynchronous Advantage Actor Critic) in 9 out of 14 tasks and FARAQL outperforms AQL (Asynchronous n-step Q-Learning) in 9 out of 13 tasks.