MAD3PG: A Framework for Multi-Agent Deep Denoising Diffusion Policy Gradient Optimization

doi:10.1016/j.inffus.2025.104026

MAD3PG: A Framework for Multi-Agent Deep Denoising Diffusion Policy Gradient Optimization

2025-12-03

0

PRE

AI

DOI：10.1016/j.inffus.2025.104026

Original

Original request for help

Save

Abstract

En 中文

• Proposes MAD3PG, the first framework using denoising diffusion models for multi-agent value distribution estimation, and provides corresponding theoretical analysis. • Introduces a K-repeated sampling strategy with temporal-difference targets to enable efficient training of diffusion models in online reinforcement learning. • Demonstrates superior robustness and efficiency over MADDPG-type algorithms through extensive experiments in MPE and MuJoCo environments.

MAD3PG: A Framework for Multi-Agent Deep Denoising Diffusion Policy Gradient Optimization

Abstract

Journal

Organization