Generating ANFISs Through Rule Interpolation: An Initial Investigation

The success of ANFIS (Adaptive-Network-based Fuzzy Inference System) mainly owes to the ability of producing nonlinear approximation via extracting effective fuzzy rules from massive training data. In certain practical problems where there is a lack of training data, however, it is difficult or even impossible to train an effective ANFIS model covering the entire problem domain. In this paper, a new ANFIS interpolation technique is proposed in an effort to implement Takagi-Sugeno fuzzy regression under such situations. It works by interpolating a group of fuzzy rules with the assistance of existing ANFISs in the neighbourhood. The proposed approach firstly constructs a rule dictionary by extracting rules from the neighbouring ANFISs, then an intermediate ANFIS is generated by exploiting the local linear embedding algorithm, and finally the resulting intermediate ANFIS is utilised as an initial ANFIS for further fine-tuning. Experimental results on both synthetic and real world data demonstrate the effectiveness of the proposed technique.