Synthetic Benchmarks for Genetic Improvement

Abstract

Genetic improvement (GI) uses automated search to find improved versions of existing software. If over the years the potential of many GI approaches have been demonstrated, the intrinsic cost of evaluating real-world software makes comparing these approaches in large-scale meta-analyses very expensive. We propose and describe a method to construct synthetic GI benchmarks, to circumvent this bottleneck and enable much faster quality assessment of GI approaches.