Virtual Genetic Algorithms: First Results

An important goal of the theory of genetic algorithms is to build models that predict how well genetic algorithms are expected to perform on a given fitness landscape (i.e., a given combination of representation, fitness function, and set of genetic operators). This paper describes the design of a software tool called a virtual genetic algorithm (VGA) that predicts the behavior of a genetic algorithm. The VGA operates like a genetic algorithm except that evaluations of individuals are based on empirically derived statistical fitness models. Because it by-passes the evaluation process, the VGA can be executed in a fraction of the time of the GA that it models, allowing multiple exploratory runs that produce average-case, best-case and worst-case predictions. We discuss ways to build the required models based on a preliminary exploration of the fitness landscape. Our initial results show that the VGA can provide a cost effective way to explore the likely performance of alternative genetic representations and operators.