Phasic Policy Gradient

Karl Cobbe, Jacob Hilton, Oleg Klimov, John Schulman

2020-09-09Reinforcement Learning

Abstract

We introduce Phasic Policy Gradient (PPG), a reinforcement learning framework which modifies traditional on-policy actor-critic methods by separating policy and value function training into distinct phases. In prior methods, one must choose between using a shared network or separate networks to represent the policy and value function. Using separate networks avoids interference between objectives, while using a shared network allows useful features to be shared. PPG is able to achieve the best of both worlds by splitting optimization into two phases, one that advances training and one that distills features. PPG also enables the value function to be more aggressively optimized with a higher level of sample reuse. Compared to PPO, we find that PPG significantly improves sample efficiency on the challenging Procgen Benchmark.

Results

Task	Dataset	Metric	Value	Model
Reinforcement Learning	ProcGen	Mean Normalized Performance	0.757	PPG
Reinforcement Learning	ProcGen	Mean Normalized Performance	0.576	PPO

Related Papers

CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning2025-07-18 VisionThink: Smart and Efficient Vision Language Model via Reinforcement Learning2025-07-17 Spectral Bellman Method: Unifying Representation and Exploration in RL2025-07-17 Aligning Humans and Robots via Reinforcement Learning from Implicit Human Feedback2025-07-17 VAR-MATH: Probing True Mathematical Reasoning in Large Language Models via Symbolic Multi-Instance Benchmarks2025-07-17 QuestA: Expanding Reasoning Capacity in LLMs via Question Augmentation2025-07-17 Inverse Reinforcement Learning Meets Large Language Model Post-Training: Basics, Advances, and Opportunities2025-07-17 Autonomous Resource Management in Microservice Systems via Reinforcement Learning2025-07-17