Modified ANFIS with Less Model Complexity for Classification Problems

A machine learning technique develops the best-fit model by adjusting weights based on learning from data. Similarly, adaptive neuro-fuzzy inference system (ANFIS) is also one of the commonly used machine learning techniques which employs training algorithm to adjust its parameters to approximate the problem under consideration. Even though, ANFIS is used in wide variety of applications including rule-based control systems, classification, and pattern matching, but ANFIS has drawback of computational complexity as it carries the problem of curse of dimensionality. This limits ANFIS to be used only with the applications having less number of inputs. Additionally, the gradient-based learning algorithm suffers from the problem of trapping in local minima. To address these drawbacks, this study reduces ANFIS architecture from five to four layers, in order to reduce model complexity. Moreover, to avoid the local minima problem in typical hybrid learning algorithm, the popular swarm-based metaheuristic algorithm Artificial Bee Colony (ABC) is used to solve the benchmark classification problems with varying input-size. The overall comparison of results and experiments show that the modified ANFIS model performed equally better as compared to standard ANFIS, but with significantly reduced trainable parameters and training computation cost. The modified ANFIS reduced the model complexity up to 93% on classification problems with large input-size.