The effect of pre-processing techniques and optimal parameters on BPNN for data classification
The architecture of artificial neural network (ANN) laid the foundation as a powerful technique in handling problems such as pattern recognition and data analysis. It’s data-driven, self-adaptive, and non-linear capabilities channel it for use in processing at high speed and ability to learn the sol...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English English |
| Published: |
2015
|
| Subjects: | |
| Online Access: | http://eprints.uthm.edu.my/1294/2/AMEER%20SALEH%20HUSSEIN%20COPYRIGHT%20DECLARATION.pdf http://eprints.uthm.edu.my/1294/1/24p%20AMEER%20SALEH%20HUSSEIN.pdf http://eprints.uthm.edu.my/1294/3/AMEER%20SALEH%20HUSSEIN%20WATERMARK.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The architecture of artificial neural network (ANN) laid the foundation as a powerful technique in handling problems such as pattern recognition and data analysis. It’s data-driven, self-adaptive, and non-linear capabilities channel it for use in processing at high speed and ability to learn the solution to a problem from a set of examples. It has been adequately applied in areas such as medical, financial, economy, and engineering. Neural network training has been a dynamic area of research, with the Multi-Layer Perceptron (MLP) trained with back propagation (BP) mostly worked on by various researchers. However, this algorithm is prone to have difficulties such as local minimum which are caused by neuron saturation in the hidden layer. Most existing approaches modify the learning model in order to add a random factor to the model which can help to overcome the tendency to sink into local minima. However, the random perturbations of the search direction and various kinds of stochastic adjustment to the current set of weights are not effective in enabling a network to escape from local minimum within a reasonable number of iterations. In this research, a performance analysis based on different activation functions; gradient descent and gradient descent with momentum, for training the BP algorithm with pre-processing techniques was executed. The Min-Max, Z-Score, and Decimal Scaling Normalization pre-processing techniques were analyzed. Results generated from the simulations reveal that the pre-processing techniques greatly increased the ANN convergence with Z-Score producing the best performance on all datasets by reaching up to 97.99%, 95.41% and 96.36% accuracy. |
|---|