A Modified Particle Swarm Optimization for Neural Networks Training on Classification Problems

This work investigates utilizing the evolutionary algorithms in training feed forward neural network. First the particle swarm optimization (PSO) algorithm is described. Then a modified version of it isintroduced. Later on multilayer perceptron network is trained to solve the circle in the square, the sphere inthe cube, and the two spirals benchmarks. The obtained results demonstrate that the PSO training algorithm can reach better classification accuracy and has more repeatability because it can escape of local minimum trap. In comparison with the back propagation (BP)algorithm, our training algorithm has more simplicityand reliability.

کلیدواژه ها:

Two Spirals Problem ، Classification ، Multilayer Perceptron ، Particle Swarm Optimization ، Backpropagation