Fitness function design for genetic algorithms in cost evaluation based problems

This work presents a class of scaling functions for genetic algorithms. These functions imply individual performance to be expressed as a set of costs. Two basic functions are obtained. Both are based in exponential functions and contain a selectivity parameter assuring an adjustable degree of discernment between individuals. The first is translation invariant while the second is both translation and scale invariant. Three examples were used to compare these scaling functions with linear scaling: an integer linear programming problem; a best path finding problem; and a best path finding problem with deceiving characteristics. In all examples, exponential based functions achieved better results than linear scaling