3 Selection Pressure and Deception

Selection methods select fitter individuals to produce new individuals. The fitness function defines a scalar value for each individual that the selection method uses to compare individuals. However, the loss of different fitness values in the population leads to the reduction of selection pressure on individuals with the same fitness. For example, if the population contains only one fitness value, then a selection method based on fitness will become random. When fitness-based diversity is low, particularly among the fitter individuals, selection and search become more random. Previous work has also linked selection pressure to code growth [Smith and Harries, 1998,Langdon and Poli, 1998a,Tackett, 1994].

Low phenotype diversity in the population can also represent another property of the population, deception. Goldberg (2002) provides a discussion of deception in context of genetic algorithms and building blocks. Langdon and Poli (1998b) describe deception in terms of the Artificial Ant problem for genetic programming. Deception, in general, refers to solutions that lead the search toward poor local optima. Deception can occur when very different solutions exist with the same fitness but their recombination leads to poor fitness. Or deception can occur when solutions that have relatively good fitness are not amenable to further improvement. Selecting these individuals would not lead to better solutions. In both cases, the selection method does not have enough information about the problem or population to avoid deceptive solutions.