Listwise ranking predictor for evolutionary neural architecture search

In evolutionary neural architecture search (ENAS), the accuracy predictors (i.e., regression models) have been successfully applied to save computational costs for the evaluation of network architectures. However, the accuracy of these predictors is largely limited by the small amount of evaluated architectures that may be difficult to obtain. Such accuracy predictors with prediction bias often lead to an inaccurate ranking, misleading the selection of ENAS. To alleviate the above limitations, we design an efficient and novel listwise ranking predictor (LRP) for ENAS to directly predict the ranking of each architecture instead of the numerical accuracy value of each architecture. Specifically, the training data is constructed by the proposed random encoding-combination (REC) strategy, which can generate substantial training data using the small number of evaluated architectures (data level). These specially constructed training data are used to train LRP, which can convert the complex regression task into a ranking task to reduce ranking bias (model level). The proposed NAS method is compared with state-of-the-art NAS methods on widely-used benchmark datasets and practical application. Experimental results demonstrate that LRP can alleviate the ranking disorder problem and outperform others in terms of both effectiveness and efficiency.