Application of genetic algorithms in production line classification systems

This work aims at simplifying the 'optimal feature set selection' step in the design of quality control systems used in the production line processes by means of genetic algorithms (GA). The goals of the work are (i) to save the engineer hours spent on examining sample distributions in the feature space to select the right set of features (ii) to extract the smallest possible feature set that returns agreeable classification accuracy, so that the system finally developed adheres to the temporal limitations imposed by the production line. The idea is to encode feature subsets directly or indirectly as chromosomes and then to assign each chromosome a fitness depending upon the number of features it uses and the testing accuracy obtained by using the classifier constructed on the chromosome's features. The paper also discusses a genetic method that discretizes continuous features while simultaneously selecting optimal features for the decision tree classifier. In the last section the performance of this method has been demonstrated by exemplifying a production line diagnosis system that has been built on the feature set prescribed by it.