Fuzzy interpretation for temporal-difference learning in anomaly detection problems

Abstract

Nowadays, information control systems based on databases develop dynamically worldwide. These systems are extensively implemented into dispatching control systems for railways, intrusion detection systems for computer security and other domains centered on big data analysis. Here, one of the main tasks is the detection and prediction of temporal anomalies, which could be a signal leading to significant (and often critical) actionable information. This paper proposes the new anomaly prevent detection technique, which allows for determining the predictive temporal structures. Presented approach is based on a hybridization of stochastic Markov reward model by using fuzzy production rules, which allow to correct Markov information based on expert knowledge about the process dynamics as well as Markov’s intuition about the probable anomaly occurring. The paper provides experiments showing the efficacy of detection and prediction. In addition, the analogy between new framework and temporal-difference learning for sequence anomaly detection is graphically illustrated.