Improving Offline RL by Blending Heuristics

TL;DR

This paper introduces Heuristic Blending (HUBL), a simple technique that improves offline reinforcement learning algorithms by combining heuristic estimates with bootstrapped values, leading to better policy performance.

Contribution

HUBL is a novel, easy-to-implement method that modifies Bellman operators in offline RL, reducing complexity and enhancing performance across multiple algorithms and benchmarks.

Findings

HUBL improves policy quality by an average of 9% across 27 datasets.

HUBL effectively combines heuristic and bootstrapped values in offline RL.

Theoretical analysis shows HUBL reduces offline RL complexity and enhances finite-sample performance.

Abstract

We propose Heuristic Blending (HUBL), a simple performance-improving technique for a broad class of offline RL algorithms based on value bootstrapping. HUBL modifies the Bellman operators used in these algorithms, partially replacing the bootstrapped values with heuristic ones that are estimated with Monte-Carlo returns. For trajectories with higher returns, HUBL relies more on the heuristic values and less on bootstrapping; otherwise, it leans more heavily on bootstrapping. HUBL is very easy to combine with many existing offline RL implementations by relabeling the offline datasets with adjusted rewards and discount factors. We derive a theory that explains HUBL's effect on offline RL as reducing offline RL's complexity and thus increasing its finite-sample performance. Furthermore, we empirically demonstrate that HUBL consistently improves the policy quality of four state-of-the-art bootstrapping-based offline RL algorithms (ATAC, CQL, TD3+BC, and IQL), by 9% on average over 27 datasets of the D4RL and Meta-World benchmarks.