summary of Meta Optimization Improving Compiler Heuristics with Machine Learning by Stephenson et al of MIT
... Notably, because many modern compiler optimizations require dealing with NP-complete problems, compiler writers must develop their own heuristics to use in the approximate solutions of these problems, and the efficacy of such heuristics is often contingent on the formulation of an appropriate priority function. ... The authors propose that their genetic programming system, Meta Optimization, is a viable resource for automatically generating both general and application specific priority functions that are probably superior to those produced by humans. The authors next provide background on how widespread and crucial priority functions are in modern compiler optimizations. ... Because of the variety of machine learning techniques available, the authors proceed to justify their choice to use genetic programming (GP) in Meta Optimization. ... They proceed to give details of the GP methodology in Meta Optimization. Meta Optimization uses a largely random initial population that is also seeded with the priority function distributed with a commercial compiler, though the authors mention that this nonrandom seed rarely affects the outcome of the system.