Reproduce “Diffusion-based Reinforcement Learning via Q-weighted Variational Policy Optimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shutong Ding, Ke Hu, Zhenhao Zhang, Kan Ren, Weinan Zhang, Jingyi Yu, Jingya Wang, Ye Shi
QVPO introduces a framework for applying diffusion models to online reinforcement learning by proposing a Q-weighted variational policy objective and an entropy regularization term. The approach achieved state-of-the-art performance and improved sample efficiency across MuJoCo continuous control benchmarks.
Key insights
- A Q-weighted Variational Lower Bound (VLO) objective can serve as a tight lower bound for the standard online Reinforcement Learning policy objective, enabling principled policy improvement for diffusion models.
- Transforming Q-values, such as using the advantage function `A(s,a) = Q(s,a)-V(s)`, effectively handles negative Q-values and guides the diffusion policy towards higher-value actions by focusing on advantageous choices.
- Explicit entropy regularization using uniformly sampled actions effectively maintains the exploration capabilities of diffusion policies, especially when computational efficiency requires fewer diffusion steps.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.