Enhancing PPO with Trajectory-Aware Hybrid Policies

TL;DR

The paper tackles the high variance and sample-inefficiency of on-policy PPO by introducing HP3O, a hybrid optimizer that reuses recent trajectories through a FIFO trajectory replay buffer and updates from a batch containing the best trajectory plus random samples. It provides theoretical policy-improvement lower bounds for HP3O and HP3O+ that account for multiple prior policies drawn from the buffer, with HP3O+ adding a value-based baseline penalty to reduce variance. Empirically, HP3O and HP3O+ demonstrate improved sample efficiency and lower variance across multiple continuous-control tasks, with runtimes comparable to PPO and often surpassing other baselines like P3O, GEPPO, and SAC. These results validate the practical utility of selectively reusing trajectory data while maintaining near on-policy stability, offering a step toward effective integration of on-/off-policy ideas.

Abstract

Proximal policy optimization (PPO) is one of the most popular state-of-the-art on-policy algorithms that has become a standard baseline in modern reinforcement learning with applications in numerous fields. Though it delivers stable performance with theoretical policy improvement guarantees, high variance, and high sample complexity still remain critical challenges in on-policy algorithms. To alleviate these issues, we propose Hybrid-Policy Proximal Policy Optimization (HP3O), which utilizes a trajectory replay buffer to make efficient use of trajectories generated by recent policies. Particularly, the buffer applies the "first in, first out" (FIFO) strategy so as to keep only the recent trajectories to attenuate the data distribution drift. A batch consisting of the trajectory with the best return and other randomly sampled ones from the buffer is used for updating the policy networks. The strategy helps the agent to improve its capability on top of the most recent best performance and in turn reduce variance empirically. We theoretically construct the policy improvement guarantees for the proposed algorithm. HP3O is validated and compared against several baseline algorithms using multiple continuous control environments. Our code is available here.

Figures (13)

• Figure 1: Schematic diagram of HP3O/HP3O+: (left side) the trajectory replay buffer takes a "first in, first out" (FIFO) strategy to keep only recent trajectories; batch consisting of the trajectory with the best return ($\tau^*$) and other randomly sampled ones from the buffer are used for updating the actor/critic networks (off-policy approach); (right side) model updating still follows the on-policy PPO method, hence, hybrid-policy PPO (HP3O); for HP3O+, $\tau^*$ is also used to update the advantage function
• Figure 2: Training curves (over 1M steps) on continuous control benchmarks. HP3O+ (black) performs consistently across all tasks and is comparable to or outperforming other baseline methods.
• Figure 3: Comparison of Normalized Standard Deviation and Runtime for 1 million steps.
• Figure 4: Explained Variance for HalfCheetah for PPO and HP3O
• Figure 5: Comparison of HP3O and PPO training curves across different environments. (a) shows the performance on Cartpole, while (b) shows the performance on Humanoid.

Theorems & Definitions (24)

• Lemma 1
• Remark 1
• Lemma 2
• Remark 2
• Theorem 1
• Remark 3
• Lemma 3
• Theorem 2
• Remark 4
• Remark 5