| With the increasing development of machine learning technologies,there is a growing demand for rapid application of machine learning systems to real-world scenarios.However,traditional machine learning model architectures and parameters need to be designed by humans,which consume a lot of human resources.Automatic machine learning(Auto ML)is a technical solution that trades arithmetic power for human power by parameterizing the model structure into a hyperparametric space and exploring the optimal solution through a large number of attempts by the program.Most of the existing automatic machine learning frameworks use traditional optimization algorithms,such as Bayesian,evolutionary or particle swarm algorithms,which require a complete iteration of the model in each attempt and are costly,and thus there is an urgent need to improve the performance and efficiency of the optimization algorithms.To address the above problems,this paper proposes to apply the BOHB(Bayesian optimization HyperBand)algorithm,which combines Bayesian optimization with multi-fidelity optimization in automatic machine learning scenarios,and improves the search efficiency of the BOHB algorithm by improving the TPE(Tree-structure Parzen estimate’)sampler of the BOHB algorithm,which solves the problems such as the inability of TPE to model the correlation between variables.In this paper,we design and implement an automatic machine learning framework based on the improved BOHB algorithm,this framework includes five main parts:storage layer,algorithm component layer,workflow layer,hyperparameter optimization layer,and visualization layer.Users can define the dataset and hyperparameter space in the code,submit optimization tasks,view and analyze the results through the front-end web page after the model optimization is completed.Through comparison experiments and system tests,it is proved that:the optimization performance and efficiency of improved BOHB algorithm are higher than other algorithms;the automatic machine learning framework is fully functional and has high robustness,scalability and practically valuable. |
