Multi-Layer Perceptron Training by Genetic Algorithms

In this paper, the authors are presenting one of the ways to train Artificial Neural Networks (ANN) using a predefined set of weights as biases as input parameters, generated by the Genetic Algorithm (GA). This approach solves the problem of hyperparameter tuning for ANN, which is an NP-hard space search problem and will be further explained in the paper. The genetic algorithm generates a population of potential solutions in each iteration and then after a series of solutions variables modification (crossover, mutation, etc.) ranks them based on their fitness values. The algorithm itself is tested on a standard Multi-layer Perceptron (MLP) artificial neural network and results are similar compared to other techniques of training.