FECS: An efficiency based learning classifier system applied to an industrial production process
Stefan Sette · AIP conference proceedings · 2001
The application of Genetic Algoritms (GA) in Rule Based Machine Learning (RBML) results in Genetic Based Machine Learning (GBML). One of the first GBML implementations is the Learning Classifier System (LCS) defined by Goldberg [4]. Learning is done by the so called Bucket Brigade Algorithm (BBA) which assigns a strength (payoff value) to each classifier based upon its interaction with the environment. However, the BBA has also several severe limitations and can therefore only be successfully applied to simple (discrete) problems. This paper introduces a new classifier system FECS based upon (several) classifier efficiency parameters. These parameters allow for a continuous (long term memory) evaluation of the classifiers and are also used to modify reward values based upon the average efficiency of the current population. Moreover, a new anticipatory mutation operator allows classifiers to mutate towards a probable better accuracy or generalism. The efficiency of FECS is demonstrated by applying it to an industrial production process. The resulting classifier set is shown to have an accuracy of 94%.