An Efficient Contesting Procedure for AutoML Optimization

Automated Machine Learning (AutoML) frameworks are designed to select the optimal combination of operators and hyperparameters. Classical AutoML-based Bayesian Optimization (BO) approaches often integrate all operator search spaces into a single space. However, disadvantage this history-based strategy is that it can be less robust when initialized randomly than optimizing each algorithm independently. To overcome issue, novel contesting procedure algorithm, Divide And Conquer (DACOpt), proposed make AutoML more robust. DACOpt partitions space reasonable number sub-spaces based on similarity budget constraints. Furthermore, throughout optimization process, allocates resources sub-space ensure (1) areas covered (2) assigned most promising sub-space. Two extensive sets experiments 117 benchmark datasets demonstrate achieves significantly better results in 36% datasets: 5% compared TPOT, 8% - AutoSklearn, 15% H20 18% ATM.