Training ANFIS Using the Enhanced Bees Algorithm and Least Squares Estimation. Issue 2 (3rd April 2017)

• Record Type: Journal Article
• Authors: Marzi, Hosein; Haj Darwish, Ahmed; Helfawi, Humam
• Is Part Of: Intelligent automation & soft computing. Volume 23:Issue 2(2017)
• Page Start: 227
• Page End: 234
• Publication Date: 2017-04-03
• Subjects: Bees Algorithm -- ANFIS -- Fuzzy Systems -- Hybrid Learning; Artificial intelligence -- Periodicals; Intelligent control systems -- Periodicals; 003.5
• DOI: 10.1080/10798587.2016.1196880
• Languages: English
• ISSNs: 1079-8587

Abstract

Abstract: This paper presents the result of research in developing a novel training model for Adaptive Neuro-Fuzzy Inference Systems (ANFIS). ANFIS integrates the learning ability of Artificial Neural Networks with the Takagi-Sugeno Fuzzy Inference System to approximate nonlinear functions. Therefore, it is considered as a Universal Estimator. The original algorithm used in ANFIS training process has a hybrid model that uses Steepest Decent Derivative; therefore, it inherits low convergence rate and local minima during training. In this study, a training algorithm is proposed that combines Bees Algorithm (BA) and Least Square Estimation (LSE) (BA-LSE). The local and global exploration of BA as integrates with the best-fit solution of the LSE improves current shortcomings of ANFIS training process. The proposed training algorithm is examined under three different scenarios of function approximation, time series prediction, and classification experiments in order to verify the promising improvements in the training process of ANFIS. The experimental results validate high generalization capabilities of the BA-LSE training algorithm in comparison to the original hybrid training model of ANFIS. The new training model also enhances local minima avoidance and has high convergence rate.