Research on Anomaly Detection Data Mining Algorithm Based on Reinforcement Learning

This paper designs a reinforcement learning framework whose state space is composed of characteristic values of samples, whose action space is defined as normal or abnormal. Through experimental Q-learning algorithm, agents are able to adjust their decision strategies automatically and learn how to identify anomalies in large amounts of data. Part of the thesis uses KDDCUP99 and NSL-KDD data sets to verify. Experimental results show that this algorithm has better performance than traditional SVM and DT. with respect to detection, recognition rate, and F1 value. On KDDCUP99 data set, the detection accuracy was 96.5%, the recall rate was 94.3%, and F1 was 95.4%. After that, the accuracy of training is improved 5% and the training time is reduced by 20%. It is proved by experiments that Q-learning algorithm based on QRS is superior in accuracy and robustness, and it is able to adapt to changes of data under dynamic environment.