Information Theory Based Learning Algorithms And Applications For Large-scale Fuzzy Cognitive Maps

Fuzzy cognitive maps(FCMs),combining fuzzy logic and neural networks,are inference networks.Since various automated FCM learning algorithms were proposed,FCMs have been widely used to model and simulate dynamic systems.However,with the large continuous search space,most of these methods were only applied to learn small-scale FCMs and the learned maps obtained by such automated learning methods are usually much denser than those constructed by experts.The main task of the learning algorithms is to learn the connection weighted matrix from observed response sequences which requires the learning methods not only to determine the edges between concepts,but also to optimize the edge weights.In this thesis,we mainly focus on the FCM learning problems,and the major work can be summarized as follows,1.A decision making trial and evaluation laboratory(DEMATEL)based genetic algorithm is proposed to learn FCMs.In the proposed algorithm,the DEMATEL method is used as a directed neighborhood search operator to steer the search to the right direction in the objective space,which can make the search jump out of local optima.Experimental results on both synthetic and real life data demonstrate the efficiency of the proposed algorithm.The comparison with existing learning algorithms shows that the proposed algorithm can learn FCMs with higher accuracy without expert knowledge.2.A mutual information(MI)based two-phase memetic algorithm(MA)is proposed for learning large-scale FCMs,and applied to solve the gene regulatory network(GRN)reconstruction problem.In the proposed method,the first phase is oriented to determine the existence of links between concepts by MI,which can reduce the search space significantly for MA,and then MA is used to optimize the edge weights in the second phase.Experiments on both synthetic and real-life data demonstrate that the proposed method can learn large-scale FCMs with an excellent performance.The application for the GRN reconstruction problem expands the application of FCM to the field of bioinformatics.3.An information theory based two-phase decomposed parallel memetic algorithm for large-scale FCM learning is proposed,and also applied to the GRN reconstruction problem.In the proposed method,the first phase is used to determine the existence of edges between concepts by the information theory,which can reduce the search space significantly,and then the second phase is oriented to optimize the edge weights from decomposed observed response sequences in parallel,where the small search space and the parallelization method substantially speed up the learning algorithm so that it is able to learn large-scale FCMs.Experiments on synthetic data with various scales and the application on the benchmark datasets DREAM3 and DREAM4 for the GRN reconstruction problem demonstrate that the proposed method is able to learn large-scale FCMs effectively.The comparison with existing learning algorithms shows that the proposed algorithm has excellent performance.