POOL-WISE CROSSOVER IN GENETIC ALGORITHMS : AN INFORMATION-THEORETIC VIEW

Several recent studies have examined the effects of replacing the pair-wise recombination operators often used by genetic algorithms with pool-wise recombination operators. In their simplest form, poolwise operators create probabilistic models that represent each parameter independently. Instead of using traditional crossover operators, these models are sampled to generate a new population. Although these simple probabilistic models are competitive with pair-wise recombination operators on many benchmark problems, pool-wise operators may not perform as well on problems which have a high degree of interdependence between the parameters. We wish to discover whether this is a fundamental limitation of pool-wise crossover or whether the poor performance is due to the simple probabilistic models which are often used. This paper presents a crossover operator, based on information theory, which explicitly captures pair-wise interdependence between parameters. By sampling this probabilistic model to generate the next population, promising results are obtained in comparison to both simple pool-wise and pair-wise operators.