Special issue on advances in Learning Classifier Systems

Abstract

Learning Classifier Systems (LCS) constitute a uniquely adaptable class of learning framework existing at the intersection of machine learning and evolutionary computation. Fundamentally, an LCS combines genetic search with an appropriate learning strategy to evolve a rule set which collectively describes a temporal or spatial problem. Since their conceptualization, a variety of algorithmic architectures and mechanisms have been introduced with the objective of improved functionality or application to new problem domains. As technology advances, the power and versatility of LCS algorithms makes them increasingly appealing. Since 1992, the International Workshop on Learning Classifier Systems (IWLCS) has sought to promote the ongoing development of LCS architecture, application, and theory. Beginning in 2003, the IWLCS has been held annually, in association with the Genetic and Evolutionary Computation Conference (GECCO). This special issue includes original works as well as selected and revised papers from the 13th and 14th IWLCS, held in Portland, USA, during GECCO 2010, and Dublin, IR, during GECCO 2011. Over the last few years, there has been tremendous progress in (1) the systematic design of novel LCSs that can deal with additional problem topologies, (2) theoretical analyses incorporated to better understand and improve existing systems and (3) applications within important domains. The works collected in this special issue extend this trend, illustrating the maturity of LCSs and their applicability to hard real-world problems that currently elude solution. In Production System Rules as Protein Complexes from Genetic Regulatory Networks: An Initial Study, Bull presents a new type of production system rules, by introducing a new, indirect encoding that views rules as protein complexes produced by the temporal behavior of an artificial genetic regulatory network. The experimental analysis shows that the new innovative approach has potential, and opens up different lines for future research. Taking a look back to the origins of LCSs, in Risk Neutrality in Learning Classifier Systems Smith develops a novel model of risk-neutral reinforcement learning in a traditional Bucket Brigade credit-allocation market which is adjusted by a genetic algorithm. The work revises two important aspects in LCSs: default hierarchies and long chains of coupled classifiers. The new model is extensively discussed, reviving some ideas present in the first LCSs proposals by Holland. In Analysing BioHEL Using Challenging Boolean Functions, the issue moves to different systematic analysis of current LCSs that enable a better understanding of the systems and give light to future improvements. In this paper, Franco et al. examine how BioHEL, a hierarchical evolutionary learning system specifically designed to cope with large-scale datasets, behaves on a family of Boolean functions whose complexity is varied along different dimensions. The experimental analysis shows the D. Loiacono Politecnico di Milano, P.zza Leonardo Da Vinci, 32, 20133 Milan, Italy e-mail: loiacono@elet.polimi.it