An Evolutionary Neural Architecture Search Method Accelerated by Multi-Fidelity Evaluation and Genetic Decision Controller

Deep neural networks (DNNs) are now widely used in numerous fields. However, manually design-ing DNNs is labor-intensive and requires expert knowledge. Evolutionary neural architecture search (ENAS) is an efficient method. Nevertheless, ENAS requires full-epoch training for each candidate architecture to determine fitness values, leading to high computational costs. To accelerate the search process and reduce resource consumption, this paper proposes MFGENAS, an accelerated ENAS method. MFGENAS is implemented within the NSGA-II framework and employs two key acceleration strategies: multi-fidelity evaluation and a genetic decision controller. We demonstrate through experi-mental analysis that classification-based prediction is significantly more effective than regression-based prediction in estimating architecture performance. To reduce the need for expensive evaluations, we introduce a genetic decision controller to evaluate the quality of generated offspring. This process is treated as a classification task: if the controller predicts that the offspring will outperform the parent, the offspring is retained. For this purpose, we adopt a kernel extreme learning machine optimized by an improved polar lights optimizer as the genetic decision controller. Comprehensive experiments vali-date MFGENAS’s performance. On the CIFAR-10 dataset, MFGENAS discovers an architecture with an error rate of 2.39% using only 0.3 GPU days. In contrast, the classical AE-CNN requires 27 GPU days to achieve an architecture with a higher error rate of 4.30%. On the CIFAR-100 dataset, MFGENAS discovers an architecture with an error rate of 16.42% in just 0.3 GPU days, whereas E2EPP requires 8.5 GPU days to reach an error rate of 22.02%.