Off-Policy Deep Reinforcement Learning without Exploration

TL;DR

This paper introduces batch-constrained reinforcement learning, enabling effective off-policy deep RL from fixed datasets by restricting actions to mitigate extrapolation errors, demonstrated on continuous control tasks.

Contribution

The paper proposes a novel batch-constrained RL algorithm that learns from fixed datasets, overcoming limitations of standard off-policy methods like DQN and DDPG.

Findings

Effective learning from fixed batch data in continuous control tasks

Outperforms standard off-policy algorithms in fixed data settings

Demonstrates robustness to extrapolation errors

Abstract

Many practical applications of reinforcement learning constrain agents to learn from a fixed batch of data which has already been gathered, without offering further possibility for data collection. In this paper, we demonstrate that due to errors introduced by extrapolation, standard off-policy deep reinforcement learning algorithms, such as DQN and DDPG, are incapable of learning with data uncorrelated to the distribution under the current policy, making them ineffective for this fixed batch setting. We introduce a novel class of off-policy algorithms, batch-constrained reinforcement learning, which restricts the action space in order to force the agent towards behaving close to on-policy with respect to a subset of the given data. We present the first continuous control deep reinforcement learning algorithm which can learn effectively from arbitrary, fixed batch data, and empirically demonstrate the quality of its behavior in several tasks.