Comprehensive-Forecast Multiobjective Genetic Programming for Neural Architecture Search
提出一种综合预测多目标遗传规划方法（CFMOGP-NAS），通过集成多种回归模型和混合训练策略，在保持精度的前提下将搜索时间减少50%，提升了神经架构搜索的效率。
Neural Architecture Search (NAS) requires global topological exploration and is hence time consuming. To address this challenge, we propose the comprehensive-forecast multiobjective genetic programming for NAS, or CFMOGP-NAS for short. By integrating the strengths of various regression models and synthesizing the forecast of multiple candidates, the accuracy and robustness of architecture predictions are enhanced. The resultant algorithm this way incorporates a strategy of a mixture of complete and partial training, which balances cost and accuracy of evaluation. To also balance the population diversity, we develop a regularized tournament scheme for genetic programming. Experimental studies show that CFMOGP-NAS achieves a 50% reduction in search time without sacrificing accuracy, and verify that it substantially improves efficacy and efficiency compared with the state-of-the-art NAS methods.