An Industry-based Development of the Learning Classifier System Technique

This paper describes the development of an Industrial Learning Classifier System for application in the steel industry. The real domain problem was the prediction and diagnosis of product quality issues in a Steel Hot Strip Mill. The properties of the data from this environment include multimodality (several optima), poor separation between fault levels and high dimensionality (many parameters). The method to develop the Learning Classifier System technique, based on deterministic simulated data, is presented. The advances made in the technique, which enhance its functionality in this type of industrial environment, are given. The novel methods developed are core to the Learning Classifier System technique and are not ‘fixes’ for given problems. They address the fitness measure, encoding alphabet, population scope, phases of training, genetic operators, life limits and removal of taxation schemes. These improvements allow the industrial LCS to function correctly in the simulated domain. Encouraging results from diagnosis of real data are presented; however, further work is needed for greater accuracy and to allow the prediction function to be used on-line. Learning Classifier Systems represent a potentially useful tool that combines the transparency of symbolic approaches (such as Decision Trees) with the learning ability of connectionist approaches (such as Artificial Neural Networks) to machine learning.