Viability Evolution: Elimination and Extinction in Evolutionary Computation

Current mainstream Evolutionary Algorithms (EA) are based on the concept of selection, encapsulated in the definition of a fitness function. Besides selection, however, the natural evolutionary process includes a phenomenon of elimination, which is linked to the ideas of viability and contingency. In this paper, we show how elimination can be modeled and integrated with selection to give rise to a new evolutionary scheme that we call Selection-Elimination EA (SE-EA). Comparing conventional EA to the newly defined SE-EA we show that SE-EA can exploit naturally some prior information about the problem at hand that is seldom, if at all, exploited by conventional EA, while avoiding the assumption of knowledge that EA based on the fitness function require but that is usually not really available. We discuss the fact that the introduction of elimination in Evolutionary Computation gives rise to the possibility of obtaining a multi-level evolutionary process that includes as a central component the phenomenon of extinction. We suggest the interpretation of an evolutionary process that includes elimination, in terms of an error-driven process and derive from it a new appreciation of the role of the interaction with the environment in the determination of the outcome of the process and of the possibility of achieving an openended evolution. The working of the SE-EA are illustrated with two examples that model its application to multiobjective engineering problems.