ABSTRACT
We have introduced the concept of Partial Evaluation (PE) in Genetic Algorithms (GAs) to deal with costly fitness functions. A GAs can be costly for many reasons and the problems are highlighted when a user tries to evaluate the solutions presented by the GAs, or when too much time or resources are required to evaluate the objective functions. Some authors try to solve these problems by using small populations, but often this is not a good solution. PE is intended to evaluate part of the generation or even parts of the individuals. We propose some strategies for solving this problem. Most of the strategies estimate the fitness of individuals who were not completely evaluated.
This contribution discusses a neural network approach to PE. We investigate the relationships between the fitness of an individual, his parents, and the way in which the crossover operation takes place. The results obtained prove to us that the PE is valid.
