Research On Feature Selection Algorithm Based On Reinforcement Learning

With the rapid development of technologies such as the Internet,the Internet of Things,and sensors,a large amount of data has been produced in production and life,and people want to excavate valuable information from these data.However,many of them show the characteristics of large sample size and high feature dimension,which undoubtedly makes it more difficult to excavate data.In view of the above problems,researchers often delete unrelated and redundant feature information by using feature selection,so as to reduce the data dimension,noise interference,and the complexity of algorithm,which make the model more simple and easy to understand.Feature selection has become a research hotspot in data mining,artificial intelligence,fault diagnosis and other fields.There are some shortcomings in traditional feature selection algorithm,such as the low accuracy of selected feature subset or larger size of the selected feature subset when performing the classification task.In view of these shortcomings,the paper has proposed a feature selection algorithm which based on encapsulated feature selection model,combined with reinforcement learning theory and using income(reward)for autonomous decision-making.Compared with the traditional feature selection algorithm on UCI dataset,the experimental results show that this algorithm can select a better feature subset,which proves the feasibility and effectiveness of algorithm.The main research contents of this paper include the following two points:(1)Based on the research of recent feature selection methods and the problems which exists,this paper combines the feature selection process with the reinforcement learning training process and proposes a new feature selection algorithm model.When performs feature selection,the agent in reinforcement learning performs training search of the feature subset through a “trial and error” method,and the feature subset is adjusted according to the feedback income of the feature subset.Finally,Agent selects the action sequence that obtains the maximum income as the optimal strategy,and obtains the result of feature selection according to the optimal strategy.Experiments show that the feature subset selected by this algorithm is better than the traditional algorithm in classification accuracy.(2)Through comparison experiments,the feasibility and effectiveness of feature selection algorithm based on reinforcement learning proposed in this paper is verified,but compared with the traditional feature selection method,the effect of feature dimension reduction and classification accuracy is slightly increased,and the time of execution of the algorithm is longer.In order to solve the problems,the algorithm was further improved,and the information theory and correlation analysis theory is introduced as "guidance experience" in the process of Agent search feature subset training,that is,in the feature subset search process,the features with higher information entropy values are preferentially added;and one of a pair of features with high Pearson correlation coefficient in the feature subset is preferentially deleted..Experiments show that the improved algorithm reduces the feature dimension compared with the original algorithm,improves the classification accuracy and shortens the algorithm execution time.