| Popis: |
There is a complex web of interactions between optimization phases in static program compilation. Because there are many different types of optimizations, and each changes the form of the program and can impact the result of subsequent optimizations, the selection of optimizations to apply is challenging and is known as the "optimization phase ordering problem." There is a need to effectively optimize the order of the optimizations and specific optimizations used based on the statistics and other features of the program to gain the most benefit. In this work we propose the use of evolved neural networks to intelligently choose which optimizations are applied and in what order, to a method or a program as a whole, based on its features. In this paper we study the use of the memetic algorithm-based neuroevolutionary system called DXNN, and a genetic algorithm based neuroevolutionary system called NEAT, to evolve such neural networks. |
