Towards A Clustering Guided Rule Interpolation for ANFIS Construction

How to construct an effective ANFIS (Adaptive Network-based Fuzzy Inference System) with insufficient (sparse) training data is a challenging problem, as the rule base of such an ANFIS model will be sparse. Fuzzy rule interpolation technique enables fuzzy inference to be performed over a sparse rule base, so it is natural to introduce FRI to support the ANFIS construction. In this work, a new clustering guided rule interpolation approach is proposed for the ANFIS construction problem. Different with most existing FRI based ANFIS construction methods that commonly conduct rule interpolation at individual rule level, the proposed method makes the interpolation to be performed on a cluster level. It adopts a clustering strategy to guide the rule selection and rule weights calculation processes, ensuring the rule similarity and diversity at the same time. Particularly, the proposed approach firstly generates a rule dictionary and divides it into several clusters. Following that a cluster guided method is designed for automated selection of relevant rules from each cluster to be included in subsequent interpolation process. Then the weight for each selected rule is calculated by considering both the cluster size and the cluster distance. Experimental results against benchmark regression datasets indicate the effectiveness of the proposed approach.