2016 5th International Conference on Reliability, Infocom Technologies and Optimization (Trends and Future Directions)

14 / 2016-06-30 19:10:56

A Novel Probabilistic-PSO Based Learning Algorithm for Optimization of Neural Networks for Benchmark Problems

Particle swarm otimization,feed forward neural network

This paper proposes a novel and modified version of standard particle swarm optimization (PSO) for optimization of initial weights and biases for feed forward neural networks (FFNN) with back propagation (BP). The combination of probabilistic-PSO and FFNN greatly help in fast convergence of FFNN in classification and prediction to various benchmark problems by eliminating the disadvantage of backpropagation of getting stuck at local minima or local maxima. The proposed probabilistic-PSO differs from the standard PSO in velocity and position parameters. In velocity parameters only particle best value is utilized for guiding the particle to move towards the target in the search space, while in standard PSO both particle best and global best values are considered for deciding the new velocity of the particle. A new parameter is introduced called as the probability parameter (P0), which decides if the standard PSO is that instead of using same random number, different particles use different random numbers to fly in search space. The proposed method is used to find the initial weights and biases for FFNN with BP, once the optimum value for initial weights and biases is evaluated the FFNN is then used for classification and prediction of various neural network benchmark problems. The benchmarking databases for neural network contain various datasets from various different domains. All datasets represent realistic problems which could be called diagnosis tasks and all the datasets consist of real world data. The results for accuracy of the proposed probabilistic-PSO method are compared with existing methods.