NPENAS: Neural Predictor Guided Evolution for Neural Architecture Search

Neural architecture search (NAS) is a promising method for automatically design neural architectures. NAS adopts a search strategy to effectively explore predefined search space to find architecture in search space with outstanding performance with minimum searching cost. Bayesian optimization and evolutionary algorithms are two commonly used search strategies, but they suffer from computationally expensive, challenge to implement, and inefficient exploration ability. In this paper, we propose to combine evolutionary algorithm with a neural predictor that can significantly improve the exploration ability of evolutionary algorithms for NAS. We designed a graph-based neural predictor with uncertainty estimation and adopted an acquisition function defined from this neural predictor to guide the evolutionary algorithm to efficiently explore the search space. We further designed a graph-based neural predictor to guide evolutionary algorithm to find potentially superior offspring and speed up the evolutionary procedure. The above two algorithms are denoted as NPENAS-BO and NPENAS-NP, respectively. Experiments illustrate our proposed algorithms outperform many recently proposed NAS algorithms on closed and open-domain search spaces with both search speed and searched architecture's performance. On closed domain search space NAS-Bench-101, with only 150 evaluated architectures, NPENAS-BO achieves a mean test error 5.9%, and NPENAS-NP reaches 5.86%, which is close to the best performance of 5.77% in the ORACLE setting. On the open domain search space DARTS, with only 2.5 GPU days, NPENAS-BO found an architecture that achieves the best test error 2.52% on CIFAR-10. With only 1.8 GPU days, NPENAS-NP finds an architecture that achieves state-of-the-art performance 2.44%. The code is publicly available at \url{https://github.com/auroua/NPENASv1}.