Analysing bioHEL using challenging boolean functions
María A. Franco, Natalio Krasnogor, Jaume Bacardit · 2010
In this work we present an exhaustive empirical analysis of the Pittsburgh-style BioHEL system using a broad set of variants of the well-known k-DNF boolean function. These functions present a broad set of possible challenges for most machine learning techniques such as varying degrees of rule specificity, class unbalance and niche overlap. Moreover, as the ideal solutions are known, one can easily assess if a learning system is able to find them, and how fast. Specifically, we study two aspects of BioHEL: its sensitivity to the coverage breakpoint parameter (that determines the degree of generality pressure applied by the fitness function) and the default rule policy. The results show that BioHEL is highly sensitive to the choice of coverage breakpoint (as was expected) and that using a suitable (known beforehand) default class allows the system to learn faster than using a majority class policy. Moreover, the experiments indicate that BioHEL scalability depends directly on both k (the specificity of the rules) and the number of DNF terms in the problem.