SAPG: Split and Aggregate Policy Gradients

TL;DR

SAPG is a novel on-policy reinforcement learning algorithm that efficiently utilizes large-scale parallel environments by splitting and aggregating data, outperforming traditional methods like PPO in complex tasks.

Contribution

Introduces SAPG, a new on-policy RL method that leverages environment splitting and importance sampling to improve performance with large-scale parallel environments.

Findings

SAPG outperforms PPO in various challenging environments.

Performance of PPO saturates with increasing environment parallelization.

SAPG effectively scales with large numbers of parallel environments.

Abstract

Despite extreme sample inefficiency, on-policy reinforcement learning, aka policy gradients, has become a fundamental tool in decision-making problems. With the recent advances in GPU-driven simulation, the ability to collect large amounts of data for RL training has scaled exponentially. However, we show that current RL methods, e.g. PPO, fail to ingest the benefit of parallelized environments beyond a certain point and their performance saturates. To address this, we propose a new on-policy RL algorithm that can effectively leverage large-scale environments by splitting them into chunks and fusing them back together via importance sampling. Our algorithm, termed SAPG, shows significantly higher performance across a variety of challenging environments where vanilla PPO and other strong baselines fail to achieve high performance. Website at https://sapg-rl.github.io/