Automatically Designing Novel Evolutionary Architectures

Abstract

Evolutionary algorithms typically follow very rigid generational cycles. These cycles are simple to implement, analyze, and use, but they pose severe constraints on design space. Consequently, typical implementations can be differentiated only across a limited range of parameters and paradigms. While parameter tuning and human expertise can help determine which algorithmic configuration may be best on a given problem class, it is likely that there exist even better optimizers outside of this traditional design space. Previous work has shown that directed graphs can be used to automatically design evolutionary cycles which outperform traditional configurations on specific benchmarks. In this work, we present a novel representation and associated variation operator for evolving such graphs inspired by techniques used for neuro-evolution, demonstrate that this can significantly improve performance and result in more complex structures, and show evidence suggesting that, for some problems, traditional cycles are not even locally optimal. We further validate the effectiveness of our algorithm against benchmarks from the IOHprofiler suite, achieving competitive results.