New crossover operators for multiple subset selection tasks

We have introduced two crossover operators, MMX-BLX and MMX-BLX, for simultaneously solving multiple feature/subset selection problems where the features may have numeric attributes and the subset sizes are not predefined. These operators differ on the level of exploration and exploitation they perform; one is designed to produce convergence controlled mutation and the other exhibits a quasi-constant mutation rate. We illustrate the characteristic of these operators by evolving pattern detectors to distinguish alcoholics from controls using their visually evoked response potentials (VERPs). This task encapsulates two groups of subset selection problems; choosing a subset of EEG leads along with the lead-weights (features with attributes) and the other that defines the temporal pattern that characterizes the alcoholic VERPs. We observed better generalization performance from MMX-BLX. Perhaps, MMX-BLX was handicapped by not having a restart mechanism. These operators are novel and appears to hold promise for solving simultaneous feature selection problems.