The issue of deletion schemes for classifier systems has received little attention. In a standard genetic algorithm a chromosome can be evaluated (assigned a reasonable fitness) immediately. In classifier systems, however, a chromosome can only be fully evaluated after many interactions with the environment, since a chromosome may generalise over many environmental states. A new technique which protects poorly evaluated chromosomes outperforms both techniques from (Wilson, 1995) on two very different single step problems. Results indicate the XCS classifier system is able to learn single step problems for which no (or few) useful generalisations can be made over the input string, despite its drive towards accurate generalisation.
[For further information on this subject see "Toward a better understanding of rule initialization and deletion". Workshop Proceedings of the Genetic and Evolutionary Computation Conference (GECCO). ISBN 978-1-59593-698-1, pp. 2777–2780. July 2007.]