ACT: Empowering Decision Transformer with Dynamic Programming via Advantage Conditioning

TL;DR

This paper introduces ACT, a method that combines dynamic programming with a decision transformer to improve offline policy optimization by conditioning actions on advantages, resulting in more robust and effective trajectory generation.

Contribution

We propose a novel approach that integrates dynamic programming with a decision transformer, enabling advantage conditioning for improved offline policy optimization.

Findings

ACT outperforms baseline methods across various benchmarks.

It demonstrates effective trajectory stitching despite environmental stochasticity.

Ablation studies validate the importance of each component.

Abstract

Decision Transformer (DT), which employs expressive sequence modeling techniques to perform action generation, has emerged as a promising approach to offline policy optimization. However, DT generates actions conditioned on a desired future return, which is known to bear some weaknesses such as the susceptibility to environmental stochasticity. To overcome DT's weaknesses, we propose to empower DT with dynamic programming. Our method comprises three steps. First, we employ in-sample value iteration to obtain approximated value functions, which involves dynamic programming over the MDP structure. Second, we evaluate action quality in context with estimated advantages. We introduce two types of advantage estimators, IAE and GAE, which are suitable for different tasks. Third, we train an Advantage-Conditioned Transformer (ACT) to generate actions conditioned on the estimated advantages. Finally, during testing, ACT generates actions conditioned on a desired advantage. Our evaluation results validate that, by leveraging the power of dynamic programming, ACT demonstrates effective trajectory stitching and robust action generation in spite of the environmental stochasticity, outperforming baseline methods across various benchmarks. Additionally, we conduct an in-depth analysis of ACT's various design choices through ablation studies. Our code is available at https://github.com/LAMDA-RL/ACT.