Parallel Genetic Learning of Fuzzy Cognitive Maps

Fuzzy Cognitive Maps (FCMs), which were introduced by Bart Kosko in 1986, are a powerful modeling technique for dynamic systems. Recently introduced automated learning method, based on real-coded genetic algorithm (RCGA), allows for establishing high-quality FCMs from historical data. The current bottleneck of this method is its scalability, which originates from large continuous search space and computational complexity of genetic optimization. To this end, the goal of this project is to propose a solution to the scalability problem. The parallel nature of genetic algorithms suggests parallel processing as the natural route to explore. In this project, we use one of the parallelization approaches to genetic algorithms, namely global single-population master-slave method, to introduce a new parallelized FCMs learning method. We investigate different hardware architectures and its influence on the computational time. The experimental studies elaborate on the quality of the proposed learning method in application to large FCMs, i.e. consisted of several dozens of concepts. Potential follow-up of this work will concern application of this method to real-life systems. Parallel Genetic Learning of Fuzzy Cognitive Maps Wojciech Stach Department of Electrical & Computer Engineering University of Alberta, Edmonton, AB T6G 2V4, Canada [email protected] Final Report for IEEE-CIS Walter Karplus Summer Research Grant, 2006