Leveraging Zero-Shot Reinforcement Learning for Supercompiler Code Optimization

A reinforcement learning agent, CodeZero, can effectively optimize code by learning an optimization policy through trial-and-error interactions with a compiler environment, and then generalizing this policy to unseen programs without further training.