Optimizing Learning Weights of Back Propagation Using Flower Pollination Algorithm for Diabetes and Thyroid Data Classification

Abstract

A number of ANN methods are used, but BP is the most commonly used algorithms to train ANNs by using the gradient descent method. Two main problems which exist in BP are slow convergence and local minima. To overcome these existing problems, global search techniques are used. This research work proposed new hybrid flower pollination based back propagation HFPBP with a modified activation function and FPBP algorithm with log-sigmoid activation function. The proposed HFPBP and FPBP algorithm search within the search space first and finds the best sub-search space. The exploration method followed in the proposed HFPBP and FPBP allows it to converge to a global optimum solution with more efficiency than the standard BPNN. The results obtained from proposed algorithms are evaluated and compared on three benchmark classification datasets, Thyroid, diabetes, and glass with standard BPNN, ABCNN, and ABC-BP algorithms. The simulation results obtained from the algorithms show that the proposed algorithm performance is better in terms of lowest MSE (0.0005) and high accuracy (99.97%).