Parallel Genetic Algorithms for Tuning a Fuzzy Data Mining System

In previous work, we have described methods that we have developed for tuning a fuzzy data mining system for intrusion detection using a hierarchical genetic algorithm. Unfortunately, the genetic algorithm approach is very slow due to the computational cost of the evaluation function. In this paper, we describe parallel implementations of the genetic algorithm that were run on both a multiprocessor Unix workstation and a high performance cluster running Linux. Very little speedup was achieved on the Unix workstation because of contention for the single file system; significant speedup was achieved with the cluster in which each node had its own file system. Experimental results of our implementations based on the master-slave parallel model and multiple population parallel model are compared, and the preliminary results indicate that the multiple population model may provide a higher quality solution for a shorter amount of time. INTRODUCTION Kuok, Fu, and Wong (1998) introduced the marriage of fuzzy logic and association rule mining to address the sharp boundary problem encountered when discretizing continuous attributes for association rule mining. We have further refined these techniques for intrusion detection applications (Luo and Bridges 2000). One of the difficulties encountered in this approach, however, is defining appropriate membership functions for the fuzzy sets that serve as values of the fuzzy attributes. We have subsequently found that genetic algorithms (GAs) are effective methods for tuning the fuzzy membership functions and for feature selection (Shi 2000; Bridges and Vaughn 2000). Unfortunately, the evaluation function for the genetic algorithm requires that data mining be performed for every member of the population at every iteration of the genetic algorithm. Although, in the intrusion detection domain, the tuning process is an off-line operation, the time required for the GA is very substantial. In order to improve the performance of the GA, we have investigated parallel implementations of the genetic algorithm for data mining on both a multiprocessor Unix workstation and a high performance cluster running Linux. In the remainder of this paper, we first describe our fuzzy data mining process and the sequential genetic algorithm that we have developed for tuning