Modified Adaptive Neuro-Fuzzy Inference System Trained by Scoutless Artificial Bee Colony

Abstract

Neuro-fuzzy systems have produced high accuracy in modeling numerous real-world applications. However, the in-built computational complexity and curse of dimensionality often cease opportunities of implementations in applications with large input size. This is also true with adaptive neuro-fuzzy inference system (ANFIS) as mostly the applications in literature are with small input size. The five-layer architecture of ANFIS is modified in this paper to reduce computational cost. For effective parameters training, the popular swarm-based metaheuristic algorithm Artificial Bee Colony (ABC) algorithm is employed after modification for enhanced convergence ability. The proposed ABC variant eliminates scout bees, hence called ABC-Scoutless, outperforms standard ABC and particle swarm optimization (PSO) on benchmark test functions. The modified ANFIS trained by ABC-Scoutless performs equally better as standard ANFIS on benchmark classification problems with different input range, but with less computational cost due to reduced number of trainable parameters.