In the world of computer operating systems, managing limited resources like computational time and memory while providing good performance is a challenge. Machine learning (ML) is increasingly used to optimize decisions in compiler optimization, specifically inlining, to reduce the size of binary files. The new technique, Iterative BC-Max, tackles these challenges by training a classification policy through supervised learning instead of traditional reinforcement learning (RL) algorithms. By compiling a program corpus using a variety of baseline policies and then learning to imitate the best inlining decisions, Iterative BC-Max aims to significantly reduce the binary file size. This innovative approach shows promise for various optimization problems and future applications beyond compiler optimization.
https://research.google/blog/scalable-self-improvement-for-compiler-optimization/