Neural Architecture Search by Learning Action Space for Monte Carlo Tree Search

Linnan Wang, Saining Xie, Teng Li, Rodrigo Fonseca, Yuandong Tian

TL;DR: A new model that learns latent actions for MCTS to the application of neural architecture search

Abstract: Neural Architecture Search (NAS) has emerged as a promising technique for automatic neural network design. However, existing NAS approaches often utilize manually designed action space, which is not directly related to the performance metric to be optimized (e.g., accuracy). As a result, using manually designed action space to perform NAS often leads to sample-inefficient explorations of architectures and thus can be sub-optimal. In order to improve sample efficiency, this paper proposes Latent Action Neural Architecture Search (LaNAS) that learns actions to recursively partition the search space into good or bad regions that contain networks with concentrated performance metrics, i.e., low variance. During the search phase, as different architecture search action sequences lead to regions of different performance, the search efficiency can be significantly improved by biasing towards the good regions. On the largest NAS dataset NASBench-101, our experimental results demonstrated that LaNAS is 22x, 14.6x, 12.4x, 6.8x, 16.5x more sample-efficient than Random Search, Regularized Evolution, Monte Carlo Tree Search, Neural Architecture Optimization, and Bayesian Optimization, respectively. When applied to the open domain, LaNAS achieves 98.0% accuracy on CIFAR-10 and 75.0% top1 accuracy on ImageNet in only 803 samples, outperforming SOTA AmoebaNet with 33x fewer samples.

Keywords: MCTS, Neural Architecture Search, Search