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