Investigating Action Encodings in Recurrent Neural Networks in Reinforcement Learning

TL;DR

This paper explores how incorporating action information into recurrent neural network architectures affects reinforcement learning performance, with empirical evaluations and discussions on design choices and future challenges.

Contribution

It systematically evaluates different methods of integrating action information into RNNs for RL, highlighting their impact and identifying future research directions.

Findings

Action encoding choices significantly influence RL performance.

Empirical results show varied effectiveness of different action integration methods.

Discussion of challenges specific to RL in designing recurrent cells.

Abstract

Building and maintaining state to learn policies and value functions is critical for deploying reinforcement learning (RL) agents in the real world. Recurrent neural networks (RNNs) have become a key point of interest for the state-building problem, and several large-scale reinforcement learning agents incorporate recurrent networks. While RNNs have become a mainstay in many RL applications, many key design choices and implementation details responsible for performance improvements are often not reported. In this work, we discuss one axis on which RNN architectures can be (and have been) modified for use in RL. Specifically, we look at how action information can be incorporated into the state update function of a recurrent cell. We discuss several choices in using action information and empirically evaluate the resulting architectures on a set of illustrative domains. Finally, we discuss future work in developing recurrent cells and discuss challenges specific to the RL setting.