A Hybrid Evolutionary Neural Networks Training applied to Phonetic Classification

Abstract

In machine learning, the most challenge is the learning process of artificial neural network that aims to determine the optima set of weights and biases. In general, gradient descent methods are the most employed as training algorithm. However, this category of algorithms converges to local optima with slow convergence. For this reason, a great number of biological and swarm inspiration are developed in the literature for avoiding the shortcomings of gradient descent algorithms. Basically, genetic algorithm (GA) is inspired from Darwin theory and more recently, evolution strategies (ES) is developed. This paper proposes a new combination between multi-layer perceptron MLP and evolutionary algorithms (EA). Two algorithms of EA are exploited known as GA and ES for training strategy by optimizing the weights and biases. This improvement leads to accelerate the speed convergency and minimize the risk of getting by local optima.  The proposed methods treat the continuous speech recognition field by assessing exactly a sub-corpus of the TIMIT datasets. The experimental results shown that the ES-MLP achieves high performance compared to other algorithms including GA-MLP and Back -propagation gradient (BP) in terms of overall classification rate with 58.81%.