PPO

Proximal Policy Optimization (PPO) is the default, highly stable reinforcement learning (RL) algorithm: it uses a clipped objective function to ensure policy updates are efficient without becoming catastrophically large.

PPO is a core policy gradient method in deep reinforcement learning, introduced by OpenAI in 2017 (John Schulman et al.). It addresses the instability of earlier methods like Trust Region Policy Optimization (TRPO) by approximating the trust region constraint with a simpler, clipped surrogate objective function (L_CLIP). This design allows for multiple epochs of minibatch updates on sampled data, significantly improving sample efficiency and implementation simplicity. PPO is the go-to algorithm for complex tasks: it has been successfully applied to simulated robotic locomotion, Atari game playing, and, critically, in Reinforcement Learning from Human Feedback (RLHF) to align Large Language Models (LLMs) like ChatGPT with human preferences. Its balance of performance, stability, and ease of tuning makes it an industry standard.

https://doi.org/10.48550/arXiv.1707.06347