Program auto-tuning through population-based stochastic optimization

Abstract

Program optimization often can and need to be decoupled from the programs innate logic. In various domains, program autotuners have been developed which can automatically search for optimal program configurations using established optimization methods. OpenTuner is a general-purpose autotuner which provides a flexible search domain representation and robust optimization techniques. This thesis extends Open-Tuner with the family of population-based stochastic optimization (PBSO) algorithms to boost OpenTuner's search capability on discrete-valued tuning problems.