Output details

Evolutionary generation of neural network classifiers—An empirical comparison

Classifiers are needed for different automated pattern classification tasks in manufacturing. How to generate accurate classifiers has been a long-standing research issue. This work is an in-depth study into the evolutionary generation of multi-layer perceptron classifiers. Two generation methods were investigated, the so-called ‘wrapper’ and ‘embedded’ methods. These were compared with manual and automatic neural network generation techniques on several benchmark problems. The embedded method produced the best results, whilst the intrinsic computational costs limited the performance of the wrapper algorithm. The work definitively showed the importance of feature selection to the accuracy and compactness of the classification results.