Feed forward neural network optimization using self adaptive differential evolution for pattern classification
Shivani Bhatia, Virendra Prasad Vishwakarma · 2016
In this paper, a multilayer perceptron feed forward neural network (MPFNN) with good self-adapting and generalization ability is presented. The network is mainly used for solving the complex problems based on classification as well as regression. Most commonly, back propagation (BP) algorithm is used to train MPFNN, which suffers from the problem of local minima, over-fitting and poor convergence. To overcome the aforementioned problem, self adaptive differential algorithm is used to optimize the weights and biases of the MPFNN in the proposed approach. The property of differential evolution (DE) of global search is used to find the optimum value of weights. The mutation factor and the crossover rate are adjusted in Self adaptive DE (SDE) to maintain exploring capability for different evolving phases. Training of MPFFNN with SDE gives better results as compared to training of MPFFNN with Genetic Algorithm (GA) and DE. The experiments are performed over UCI datasets namely iris, glass, wine and ionosphere.