Research On Automatic Data Analysis Based On Deep Reinforcement Learning

With the advent of the era of big data,data analysis has become more and more difficult.Traditional data analysis techniques and methods can hardly meet the current needs of data analysis.At the same time,with the widely use of information technology in various domains,more and more fields are in urgent need of experienced data scientists.However,the experience and the quantity of data scientists can no longer meet the needs of industry development.It is particularly important to automate the data analysis in purpose of making data analysis get better and faster application in more fields.In order to achieve automatic data analysis,two fundamental problems must be solved: "Automatic algorithm selection and hyperparameters tuning for the selected model".An approach for autonomous optimization of hyperparameter and simultaneous optimization of model selection with hyperparameter has been proposed,which is based on the relevant theories and methods of deep reinforcement learning.Moreover,the solution for the problems of “unstable performance of optimization” and “the process of exploration and optimization is time-consuming” in the previous scheme has been proposed.The main contributions of this work include:(1)Automating hyperparamters tuning.According to the characteristics in the process of hyperparameters optimization,a deep neural network has been exploited for constructing an agent to replace human decision in this thesis.The parameters of the agent are learned through the technology of reinforcement learning.Thus,when a user has assigned a learning algorithm and the space of hyperparameters,the agent can automatically select a configuration of hyperparamters for a given data set.(2)Improving performance during training.To solve the problem of unstable optimization performance and to avoid falling into local optimum,an improved scheme with "Bootstrap Pool" was proposed.Through comparative experiments,it is found that this mechanism can significantly improve the stability of the optimization performance and avoid falling into local optimum.(3)Reducing training time.An improvement scheme of "Real-Prediction-Real" is proposed in order to deal with the problem of "the exploration and optimization process takes a long time" in the previous hyperparameter optimization scheme.Through comparative experiments,it is found that the method with this structure can greatly shorten the exploration and optimization time.(4)Automatic algorithm selection and hyperparameters tuning.According to the characteristics of the simultaneous optimization process of model selection and hyperparameter tuning,a method to realize the simultaneous optimization of model selection and hyperparameter based on the theory and method of deep reinforcement learning has been proposed.The experimental results show that our approach significantly outperforms other methods.