We compare two approaches in selecting neural network learning parameters and architecture. Traditionally they are found by trial and error (handcrafted) and alternatively, can be found using a genetic algorithm. Trial and error can find good solutions but the drawback is this method is time consuming and it can only try a few possible solutions while the genetic algorithm is known to be able to search for a good solution intelligently and faster with greater diversity of possible solutions. We tested the approaches on ten isolated Malay digits from 0 to 9. Three factors are compared between the two approaches: time to get a good solution; network learning convergence; and the recognition rate. Our findings show that the neural network using the genetic algorithm achieved 94% recognition rate while the handcrafted neural network achieved 95%. However, using the genetic algorithm, a good solution can be found within days while with the handcrafted method it took weeks. The network learning convergence for both approaches were relatively the same.
[1]
Robert F. Harrison,et al.
Optimization and training of feedforward neural networks by genetic algorithms
,
1991
.
[2]
John R. Koza,et al.
Genetic generation of both the weights and architecture for a neural network
,
1991,
IJCNN-91-Seattle International Joint Conference on Neural Networks.
[3]
Biing-Hwang Juang,et al.
Fundamentals of speech recognition
,
1993,
Prentice Hall signal processing series.
[4]
Riccardo,et al.
Evolution of the Topology and the Weightsof Neural Networks using GeneticProgramming with a Dual RepresentationJo
,
1997
.