Illinois Genetic Algorithms Laboratory

Llorà, X., Sastry, K., Lima, C. F., Lobo, F. G., Goldberg, D. E. (2008)
TR No.: 2008005 | Download PDF | Download PS

Abstract: This paper reviews a competent Pittsburgh LCS that automatically mines important substructures of the underlying problems and takes problems that were intractable with  first-generation Pittsburgh LCS and renders them tractable. Specifically, we propose a χ-ary extended compact classifier system  which uses (1) a competent genetic algorithm (GA) in the form of $\chi$-ary extended compact genetic algorithm, and (2) a niching method in the form restricted tournament replacement, to evolve a set of maximally accurate and maximally general rules. Besides showing that linkage exist on the multiplexer problem, and that χeCCS scales exponentially with the number  of address bits (building block size) and quadratically with the problem  size, this paper also explores non-traditional rule encodings. Gene expression encodings, such as the Karva language, can also be used to build χeCCS probabilistic models. However, results show that  the traditional ternary encoding { 0,1,#} presents a better scalability  than the gene expression inspired ones.