Effects of Weight Initialization in a Feedforward Neural Network for Classification Using a Modified Genetic Algorithm

In this paper electroencephalography (EEG) patterns are classified using a feedforward neural network trained with a modified genetic algorithm (GA). The objective is to investigate the effects of weight initialization in the neural network and to propose the best settings. Special operators like geometric ranking selection, blend-alpha crossover and non-uniform mutation are employed. For the initialization of the chromosomes the effect of the Nguyen-Widrow weight initialization and the random initialization on the training performance are compared. For the EEG corpus it is shown that the Nguyen-Widrow algorithm is more effective than the random method weight initialization when it is used with a stochastic training method. Compared to a previous study which used backpropagation as a training method, the error rate is decreased by around 10 percent if the GA is used as a training method.