A new differential evolution algorithm for cooperative fuzzy rule mining: application to anomaly detection

Fuzzy classification rule mining can be considered as a challenging optimization problem with the purpose of extracting accurate and interpretable rules. This paper deals with a Michigan cooperative approach for mining fuzzy rules. The proposed algorithm named EDE-FRMiner uses an enhanced differential evolution that evolves population of individuals; each one represents a single rule. The whole population collaborates to generate in one shot an accurate and reduced number of rules. EDE-FRMiner is an intelligent process of evolution that uses fast arithmetic operators and a new cooperative weights memory. This latter allows sharing information between individuals. In addition, it uses a new threshold based fitness function using a redefined support and confidence measures. The adaptive threshold mechanism used in the fitness function aims to adapt the miner system to problems with dynamic training data. Experiments are carried out using the NSL-kdd’99 intrusion detection data set and other data sets from the UCI repository. A comparative study with other competitive evolutionary rule based systems is performed and the results show the effectiveness of proposed algorithm.