Diagnosing Bottlenecks in Deep Q-learning Algorithms

TL;DR

This paper investigates the causes of bottlenecks in Deep Q-learning, revealing how neural network size, overfitting, and sampling strategies impact learning stability and proposing a novel sampling method for improvement.

Contribution

It introduces a unit testing framework for Q-learning, analyzes the effects of neural network architecture and sampling, and proposes a new sampling method to mitigate function approximation errors.

Findings

Large neural networks improve stability

Overfitting can be mitigated with practical techniques

A new sampling method improves performance in continuous control

Abstract

Q-learning methods represent a commonly used class of algorithms in reinforcement learning: they are generally efficient and simple, and can be combined readily with function approximators for deep reinforcement learning (RL). However, the behavior of Q-learning methods with function approximation is poorly understood, both theoretically and empirically. In this work, we aim to experimentally investigate potential issues in Q-learning, by means of a "unit testing" framework where we can utilize oracles to disentangle sources of error. Specifically, we investigate questions related to function approximation, sampling error and nonstationarity, and where available, verify if trends found in oracle settings hold true with modern deep RL methods. We find that large neural network architectures have many benefits with regards to learning stability; offer several practical compensations for overfitting; and develop a novel sampling method based on explicitly compensating for function approximation error that yields fair improvement on high-dimensional continuous control domains.