Population-based guiding for evolutionary neural architecture search

Abstract

Abstract Neural Architecture Search (NAS)—combined with biology-inspired evolutionary methods—can help discover suitable architectures tailored to a given objective. A guided evolutionary approach can enhance efficiency, aiming to accelerate the discovery of top-performing architectures within a given search space. We propose a novel algorithmic framework that implements selection, crossover, and mutation operations to generate new candidate architectures during an evolutionary Neural Architecture Search: A greedy selection operator, relying solely on model accuracy data, promotes exploitation. Incorporating architecture embeddings to further refine the mutation process enhances exploration. We introduce a guided mutation approach to steer the search toward unexplored regions of the current population. The proposed strategy, PBG (Population-Based Guiding), synergizes both explorative and exploitative methods. It substantially outperforms baseline methods such as regularized evolution by being up to three times faster on NAS-Bench-101. This combined approach not only leverages the strengths of both explorative guided mutation and exploitative greedy selection strategies, but also provides a robust and efficient framework reaching competitive performance for evolutionary Neural Architecture Search across benchmarks.