Training of Artificial Neural Network Using New Initialization Approach of Particle Swarm Optimization for Data Classification

Artificial neural network (ANN) has a wide variety of practice for the solution of problems in the area of data classification. Back propagation algorithm is a famous neural network (NN) traditional training approach. Hence, this classical training technique has many drawbacks like stuck in the local minima and maximum number of iterations required. Particle Swam Optimization (PSO) has been widely applied for the solutions of data classification problems. Population initialization is a vital factor in PSO algorithm, which considerably influences the diversity and convergence during the PSO's process. In this paper, the training of the ANN has been implemented with new initialization technique by using low discrepancies sequence, Torus termed as TO-PSO. In this paper, a detailed comparative performance analysis for the training of neural network is observed on nine benchmark data sets taken from UCI repository. The Results demonstrate that training of ANN with proposed initialization technique offer efficient and best substitute to traditional training approaches of the NN, which gives the solution of problems related to the data classification. Furthermore, the performance of TO-PSO has been compared with back propagation algorithm (BPA), standard PSO-NN and two other initialization approaches Sobol based PSO (SO-PSONN) and Halton based PSO (H-PSONN) for the training of ANN. The experimental results show that the proposed approach outperforms than BPA, traditional PSONN, SO-PSONN and H-PSONN in terms of converging speed and better accuracy Moreover, the outcomes of our work present a foresight that how the proposed initialization technique can be used as an efficient alternative to standard training approaches for the data classification problems.