Design approach to improve the artificial neural network by incorporating optimized algorithm

Abstract

Training neural networks in classification problems is a very challenging task. Many raining algorithms have been designed to boost the performance of neural networks. A popular approach is to employ a different adaptive learning rate for each weight. Most of the existing algorithms of this class are based on the use of meta-heuristics. These algorithms don t guarantee convergence to a local minimizer from any initial weights set. The algorithms designed so far converge frequently at the local minima. To alleviate this situation, this Ph.D. thesis proposes new methods that overcome these problems. newline newlineUnlike previous research in this area, though, the work focuses on handling the difficult problem constraints in a simple and effective way, without complicating the overall solution methodology. This research investigates meta-heuristic approaches for solving convergence to local minima during training of the solution algorithm that are directed to achieve desired goal. newline newlineMeta-heuristic algorithms have been successfully used in optimization problems for training Artificial Neural Network. Based on this perception, we tailored meta-heuristic algorithms for solving the problems. Among these algorithms are: Particle swarm optimization algorithm, Ant colony optimization and Bat Algorithm. newlineThey have been widely tested and applied to many difficult optimistic problems, where conventional computer science techniques are acting discontentedly. One of the newest participants in this field is the Bat algorithm, which is based on the behavior of bat echolocation.It has been proven to have great convergence properties in various reference newlinefunctions and it seems to be encouraging to manage optimistic issues. newlineIn general, the findings of the research indicate the success of our approach in handling local convergence and devising simple modified Bat algorithm in training Radial Basis function and used in a variety of benchmark datasets. In this research, we proposed BAT to train ANNs.