A Modified Extreme Learning ANFIS for Higher Dimensional Regression Problems

Abstract

Extreme learning adaptive neuro-fuzzy inference system (ELANFIS) is a new learning machine which integrates reduction of computational complexity of extreme learning machine (ELM) concept to ANFIS. ELANFIS uses Takagi–Sugeno–Kang (TSK) fuzzy inference system like ANFIS which gives accurate models. Grid partitioning is used in both ANFIS and ELANFIS which has the disadvantage of curse of dimensionality. In this paper, a modified ELANFIS using sub-clustering for input space partitioning is proposed for higher dimensional regression problems. In the proposed structure, sub-clustering is used for input space partitioning of the network. The cluster centers are used to obtain the premise parameters of the ELANFIS, such that it satisfies the constraints for obtaining distinguishable membership functions. Performance of the modified ELANFIS is compared with ANFIS and ELANFIS for real-world higher dimensional regression problems. The modified ELANFIS overcomes the curse of dimensionality with better interpretability compared to ANFIS and ELANFIS.