The problem with DDPG: understanding failures in deterministic environments with sparse rewards

TL;DR

This paper analyzes why deterministic reinforcement learning algorithms like DDPG fail in sparse reward, deterministic environments, revealing that they can get stuck in poor fixed points and providing insights for improving convergence.

Contribution

It offers a formal explanation for failures of actor-critic algorithms in sparse, deterministic settings and analyzes the underlying mechanisms causing convergence issues.

Findings

Learning can get stuck in poor fixed points.

Analysis of convergence regimes in deterministic environments.

Insights for developing better algorithms.

Abstract

In environments with continuous state and action spaces, state-of-the-art actor-critic reinforcement learning algorithms can solve very complex problems, yet can also fail in environments that seem trivial, but the reason for such failures is still poorly understood. In this paper, we contribute a formal explanation of these failures in the particular case of sparse reward and deterministic environments. First, using a very elementary control problem, we illustrate that the learning process can get stuck into a fixed point corresponding to a poor solution. Then, generalizing from the studied example, we provide a detailed analysis of the underlying mechanisms which results in a new understanding of one of the convergence regimes of these algorithms. The resulting perspective casts a new light on already existing solutions to the issues we have highlighted, and suggests other potential approaches.