Evolution of Neural Network Architecture and

Evolution of Neural Network Architecture and Weights Using Mutation Based Genetic Algorithm
A. Nadi*, S. S. Tayarani-Bathaie** and R. Safabakhsh***
* **

Amirkabir Univ. of Technol./Dept. of Comput. Eng., Tehran, Iran. Email: a.nadi@aut.ac.ir Amirkabir Univ. of Technol./Dept. of Elec. Eng., Tehran, Iran. Email: s.t.bathaie@aut.ac.ir *** Amirkabir Univ. of Technol./Dept. of Comput. Eng., Tehran, Iran. Email: safa@aut.ac.ir

Abstract— In this paper we present a new approach for evolving an optimized neural network architecture for a three layer feedforward neural network with a mutation based genetic algorithm. In this study we will optimize the weights and the network architecture simultaneously through a new presentation for the three layer feedforward neural network. The goal of the method is to find the optimal number of neurons and their appropriate weights. This optimization problem so far has been solved by looking at the general architecture of the network but we optimize the individual neurons of the hidden layer. This change results in a search space with much higher resolution and an increased speed of convergence. Evaluation of algorithm by 3 data sets reveals that this method shows a very good performance in comparison to current methods. Keywords Neural Network; MLP; Genetic Algorithm; Mutation; Architecture; Optimization

I. INTRODUCTION There may be many different Neural Networks (NN) for solving every particular problem and network designers usually face questions such as “How could we find the size of a NN?”, “Is the selected architecture an appropriate one?”, “How could one design be an optimal network?”. These questions usually lead the designer to optimization methods to find the desired network. In order to avoid the local minima encountered in most of the optimization methods, scientists tend to use random search methods such as evolutionary algorithm (EA) and Genetic algorithm (GA) to find an optimal network. Evolutionary algorithms...