An Idiosyncrasy of Time-discretization in Reinforcement Learning

TL;DR

This paper investigates how the choice of time-discretization affects reinforcement learning in continuous-time environments, highlighting an idiosyncrasy and proposing a simple modification for better alignment of return definitions.

Contribution

It identifies a specific issue with applying discrete-time RL algorithms to discretized continuous-time systems and proposes a modification to improve return consistency.

Findings

Naive application of discrete-time algorithms can misalign returns in continuous-time environments.

A simple modification can better align discrete-time returns with continuous-time definitions.

The work has practical implications for environments with stochastic or variable time discretization.

Abstract

Many reinforcement learning algorithms are built on an assumption that an agent interacts with an environment over fixed-duration, discrete time steps. However, physical systems are continuous in time, requiring a choice of time-discretization granularity when digitally controlling them. Furthermore, such systems do not wait for decisions to be made before advancing the environment state, necessitating the study of how the choice of discretization may affect a reinforcement learning algorithm. In this work, we consider the relationship between the definitions of the continuous-time and discrete-time returns. Specifically, we acknowledge an idiosyncrasy with naively applying a discrete-time algorithm to a discretized continuous-time environment, and note how a simple modification can better align the return definitions. This observation is of practical consideration when dealing with environments where time-discretization granularity is a choice, or situations where such granularity is inherently stochastic.