拡散モデルとPPOの統合による強化学習のサンプル効率と探索の向上

原文アブストラクト

Proximal Policy Optimization (PPO) is widely used in continuous control due to its robustness and stable training, yet it remains sample-inefficient in tasks with expensive interactions and high-dimensional action spaces. This paper proposes PPO-DAP (PPO with Diffusion Action Prior), a strictly on-policy framework that improves exploration quality and learning efficiency without modifying the PPO objective. PPO-DAP follows a two-stage protocol. Offline, we pretrain a conditional diffusion action prior on logged trajectories to cover the action distribution supported by the behavior policy. Online, PPO updates the actor-critic only using newly collected on-policy rollouts, while the diffusion prior is adapted around the on-policy state distribution via parameter-efficient tuning (Adapter/LoRA) over a small parameter subset. For each on-policy state, the prior generates multiple action proposals and concentrates them toward high-value regions using critic-based energy reweighting and in-denoising gradient guidance. These proposals affect the actor only through a low-weight imitation loss and an optional soft KL regularizer to the prior; importantly, PPO gradients are never backpropagated through offline logs or purely synthetic trajectories. We further analyze the method from a dual-proximal perspective and derive a one-step performance lower bound. Across eight MuJoCo continuous-control tasks under a unified online budget of 1.0M environment steps, PPO-DAP consistently improves early learning efficiency (area under the learning curve over the first 40 epochs, ALC@40) and matches or exceeds the strongest on-policy baselines in final return on 6/8 tasks, with modest overhead (1.18+/-0.04x wall-clock time and 1.05+/-0.02x peak GPU memory relative to PPO).