Learning of fuzzy cognitive maps using a niching-based multi-modal multi-agent genetic algorithm

Abstract

Fuzzy cognitive maps (FCMs) are generally applied to model and analyze complex dynamical systems. To learn the FCM weight matrix, various efficient learning algorithms have been proposed. However, those algorithms only learn one FCM from data once. Learning only one FCM is not enough for modeling and analyzing complex dynamical systems because the learned FCM may be just a local optimum. To solve this problem, we tend to learn multiple FCMs simultaneously. To this end, the FCM learning problem is modeled as a multi-modal optimization problem. So far, niching is the most adopted method to deal with multi-modal optimization. Thus, a multi-agent genetic algorithm (MAGA), which is a popular numerical optimization algorithm, is combined with current niching methods. Then, a niching-based multi-modal multi-agent genetic algorithm is proposed for learning FCM, termed as NMMMAGA-FCM. In this paper, NMMMAGA-FCM is adopted to learn several FCMs at the same time, then chooses the optimal FCM from all candidates. In the experiments, NMMMAGA-FCM is applied to learn the FCMs from synthetic data with varying sizes and densities. The experimental results show that NMMMAGA-FCM can learn FCMs with high accuracy. In addition, NMMMAGA-FCM is validated on the benchmark datasets DREAM3 and DREAM4. The experimental results show that NMMMAGA-FCM outperforms other learning algorithms obviously, which illustrates that NMMMAGA-FCM can reconstruct gene regulatory networks (GRNs) effectively.