DIAR: Diffusion-model-guided Implicit Q-learning with Adaptive Revaluation

TL;DR

DIAR is a novel offline reinforcement learning framework that uses diffusion models and adaptive revaluation to improve decision-making, robustness, and generalization in long-horizon, sparse-reward tasks.

Contribution

The paper introduces DIAR, integrating diffusion models with implicit Q-learning and adaptive revaluation for enhanced offline RL performance.

Findings

Outperforms state-of-the-art algorithms in Maze2D, AntMaze, and Kitchen tasks.

Effectively handles out-of-distribution samples and long-horizon problems.

Improves policy robustness and generalization through diverse latent trajectories.

Abstract

We propose a novel offline reinforcement learning (offline RL) approach, introducing the Diffusion-model-guided Implicit Q-learning with Adaptive Revaluation (DIAR) framework. We address two key challenges in offline RL: out-of-distribution samples and long-horizon problems. We leverage diffusion models to learn state-action sequence distributions and incorporate value functions for more balanced and adaptive decision-making. DIAR introduces an Adaptive Revaluation mechanism that dynamically adjusts decision lengths by comparing current and future state values, enabling flexible long-term decision-making. Furthermore, we address Q-value overestimation by combining Q-network learning with a value function guided by a diffusion model. The diffusion model generates diverse latent trajectories, enhancing policy robustness and generalization. As demonstrated in tasks like Maze2D, AntMaze, and Kitchen, DIAR consistently outperforms state-of-the-art algorithms in long-horizon, sparse-reward environments.