An investigation of techniques for improving the performance of a Pittsburgh approach learning classifier system
Ashley Stacey · The University of Bath Online Publications Store (The University of Bath) · 2004
Learning Classifier Systems (Holland 1978) are a Machine Learning technique in which a set of simplified production rules are discovered to solve a given problem.The system must identify patterns within the data it receives from the problem environment, enabling it to classify other, previously unseen inputs.The rules, known as classifiers, are selected and manipulated using a Genetic Algorithm (Holland 1975).This dissertation focuses on Pittsburgh approach classifier systems (Smith 1980), where many candidate rule sets compete with one another.Such systems tend to be slow since the Genetic Algorithm must work with large structures, namely entire sets of classifiers; increasing their efficiency is therefore an important research topic.We propose a technique in which the rule sets are compressed before manipulation and then re-expanded using principles from Grammatical Evolution.We then go on to explore methods of controlling bloat, a problem where rule sets grow out of control, slowing down the system and reducing the effectiveness of the search.A new algorithm is implemented, in which weak classifiers are identified and explicitly removed from the candidate solutions.This results in a considerable improvement to the system's performance on a number of Data Mining tasks.