Scheme rearrangement in a genetic algorithm: Can it help assessing latent interactions?

‘Scheme rearrangement’ was initially proposed to improve the fit and the robustness of a genetic algorithm. It was hypothesized that, in a problem comprising non linear or epistatic interactions, a greater efficiency would result from bringing closer the parameters of the interacting variables. The method was assessed, with mild success, in the specific case of the travelling salesman but not in multivariate models. It was not tested neither with real-coded parameters, though they are more adapted for identifying meaningful schemes. In the paper, we check whether scheme rearrangement provides better fit and whether it may help identifying latent interactions. These issues are studied and empirically tested with a typical problem of market segmentation: the clusterwise regression. The study supports the hypothesis that scheme rearrangement provides better solutions but not that it results from latent interactions.