Extended Evolutionary Learning of Fuzzy Cognitive Maps for the Prediction of Multivariate Time-Series

Abstract

Fuzzy cognitive maps (FCMs) is a knowledge representation tool that can be exploited for predicting multivariate time-series. FCM model represents dependencies among data variables as a directed, weighted graph of fuzzy sets (concepts). This way, FCM can be easily interpreted or constructed by experts in contrary to black box knowledge representation methods. Since FCM is a parametric model, it can be trained using historical data. So far, the genetic algorithm has been used to solely optimize the weights of FCM leaving the rest of FCM parameters to be adjusted by experts. Previous studies have shown that the genetic algorithm can be also used not only for optimizing the weights but also for optimization of FCM transformation functions. The main idea presented in this chapter is to further extend FCM evolutionary learning process. Special focus is given on fuzzyfication and transformation function optimization, applied in each concept seperately, in order to improve the efficacy of time-series prediction. The proposed extended evolutionary optimization process was evaluated in a number of real medical data gathered from the internal care unit (ICU). Comparing this approach with other known genetic-based learning algorithms, less prediction errors were observed for this dataset.