Hybrid Multi-population Evolution based on Genetic Algorithm and Regularized Evolution for Neural Architecture Search
Phanomphon Yotchon, Yutana Jewajinda · 2020
This paper presents a hybrid multi-population evolution algorithm based on genetic algorithms and regularized evolution for neural architecture search. The proposed algorithm improves the search quality by maintaining population diversity. The total population is divided into multiple tribes. In each tribe, there are multiple subpopulations with different crossover and mutation strategies. The best individuals in subpopulations are migrated between subpopulations and tribes to guide search direction toward optimum solutions. The experimental results show that the proposed algorithm performs better than traditional multi-population genetic algorithms and comparably compete with regularized evolution and other state of the art NAS algorithm on CIFAR-10 dataset.