Enhancing the harmony search algorithm for the training of multi-layer perceptron neural networks
Loading...
Date
2010
Authors
Kattan, Ali R. Mustafa
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Multi-Layer Perceptron (MLP) is a type of F~ed-Forward Artificial Neural
Networks (FFANNs) that has no lateral connections. There are basically two FFANN
supervised training paradigms: the trajectory-driven paradigm and the evolutionary
stochastic global optimization (SGO) paradigm. One of the relatively young SGO
methods is the Harmony Search (HS) algorithm, which draws its inspiration not from
biological or physical processes but from the improvisation process of musicians. HS
was reported to be competitive alternative to other SGO methods. It has been used
successfully in many applications mostly in engineering and industry.
The main aim of this work is to propose a new method for MLPs training based on
the HS algorithm. Checking the existing literature revealed that no such method exists
for any type of artificial neural network. The HS algorithm is adapted for such a task
and the performance is evaluated. Then two enhancements are introduced to achieve
better convergence condition and better performance. The potential of the new
method would be demonstrated empirically by using a set of pattern-classification
benchmarking problems. The merits of the new method are verified by comparisons
against a trajectory-driven training method, namely backpropagation, as well as an
evolutionary SGO training method, namely genetic algorithm, using the same set of
benchmarking problems. Results show that the proposed HS-based method is a
competitive FF ANN training alternative that is superior in terms of the achieved
overall recognition percentage and the overall training time in most of the
benchmarking problems used.