Genetic Algorithm with Linkage Learning

Next-generation genetic algorithms (GAs) should explore information from the problem structure whenever possible. Variable interactions can be inferred using linkage learning. Statistical linkage learning techniques were shown to improve GAs' effectiveness significantly in many problems, but may eventually report false linkages. On the other hand, empirical linkage learning (ELL) techniques discover only true variable dependencies. However, traditional ELL techniques are computationally expensive. We introduce the genetic algorithm with linkage learning (GAwLL), which discovers an empirical weighted variable interaction graph (VIGw) as a side-effect of the optimization performed by a GA, making it a no-cost ELL technique. Vertices of the VIGw represent decision variables and weights indicate the strength of the interaction between variables. The VIGw allows us to obtain new insights about the optimization problem and can be used to design genetic operators that efficiently explore the information about variable dependencies. Experiments with NK landscapes show that GAwLL is able to efficiently build the empirical VIGw. We also present an interesting machine learning application, where the VIGw represents a feature interaction network. By using GAwLL, the feature interaction network is built as a side-effect of evolutionary feature selection.