A Novel Artificial Neural Network (ANN) Using The Mayfly Algorithm for Classification

Abstract

Training of Artificial Neural Networks (ANNs) have been improved over the years using meta heuristic algorithms that introduce randomness into the training method but they might be prone to falling into a local minima in a high-dimensional space and have low convergence rate with the iterative process. To cater for the inefficiencies of training such an ANN, a novel neural network is presented in this paper using the bio-inspired algorithm of the movement and mating of the mayflies. The proposed Mayfly algorithm is explored as a means to update weights and biases of the neural network. As compared to previous meta heuristic algorithms, the proposed approach finds the global minima cost at far less number of iterations and with higher accuracy. The network proposed, which is named Mayfly Algorithm based Neural Network (MFANN) consists of an input layer, a single hidden layer of 10 neurons and an output layer. Two University of California Irvine (UCI) database sample datasets have been used as benchmark for this study, namely the Banknote Authentication (BA) and the Cryotherapy, for which the training accuracy achieved is 99.2350% and 96.6102%, whereas the Testing accuracy is 99.1247% and 90% respectively. Comparative study with grey wolf optimization neural network (GWONN) and particle swarm optimization neural network (PSONN) reveal that the proposed MFANN achieves 1-2% better accuracy with training dataset and 2% better accuracy with testing dataset.