Learning and aggregation of fuzzy cognitive maps - An evolutionary approach

Fuzzy Cognitive Maps (FCMs) are a widely used, neuro-fuzzy based qualitative approach for the modeling of dynamic systems, which allow for both static and dynamic analyses. They are capable of modeling complex systems with nonlinearities and unknown physical behaviour. FCMs describe a given system by means of concepts connected by quantified cause-effect relationships. This dissertation contributes to the subject of computer-driven generation of FCMs that can be used to perform an accurate dynamic analysis of the modeled system. The dynamic analysis provides insights into the degree of presence, and dependencies between the concepts in successive iterations of the simulation of a given FCM model. Such simulation studies could be used to analyze "what-if" scenarios in the context of decision support and to perform time series predictions. Two research directions within the framework of FCM development, which concern the learning of FCMs from historical data and an aggregation of FCMs that were proposed by multiple experts, are investigated. Several new automated computational methods for data-driven learning and aggregation of FCMs are introduced and empirically evaluated. These methods utilize real-coded genetic algorithms (RCGA)-based optimization. This choice of the optimization vehicle was motivated by their well-documented efficiency in searching large and continuous search spaces, which are inherent to our problem. Experimental evaluation demonstrates that the proposed RCGA-based learning method outperforms modern existing approaches when the dynamic analysis is considered. A novel divide and conquer-based learning strategy to improve scalability of the RCGA approach, is also proposed. This strategy is shown to be competitive or even better than solutions based on the parallelization of the underlying genetic algorithm. The RCGA-based learning method is further extended to provide improved FCMs when the number of connections of the map is known a priori. Experimental evaluation shows that the density-based learning method outperforms the generic RCGA-based approach when using a relatively accurate density estimate, and that both methods are equivalent when the estimate is inaccurate. In addition, a novel method for the aggregation of multiple input FCMs, is proposed. When compared to existing aggregation approaches, this method provides solutions that are more accurate when dynamic analysis is the objective.