Surrogate-assisted evolutionary neural architecture search with network embedding

Abstract To accelerate the performance estimation in neural architecture search, recently proposed algorithms adopt surrogate models to predict the performance of neural architectures instead of training the network from scratch. However, it is time-consuming to collect sufficient labeled architectures for surrogate model training. To enhance the capability of surrogate models using a small amount of training data, we propose a surrogate-assisted evolutionary algorithm with network embedding for neural architecture search (SAENAS-NE). Here, an unsupervised learning method is used to generate meaningful representation of each architecture and the architectures with more similar structures are closer in the embedding space, which considerably benefits the training of surrogate models. In addition, a new environmental selection based on a reference population is designed to keep diversity of the population in each generation and an infill criterion for handling the trade-off between convergence and model uncertainty is proposed for re-evaluation. Experimental results on three different NASBench and DARTS search space illustrate that network embedding makes the surrogate model achieve comparable or superior performance. The superiority of our proposed method SAENAS-NE over other state-of-the-art neural architecture algorithm has been verified in the experiments.