Evolution of Fitness Functions to Improve Heuristic Performance

Improving Metaheuristic Performance by Evolving a Variable Fitness Function

“…These are known good values from previous work [1] and this previous work also shows that the sensitivity to parameters is low.…”

“…First presented in [1], the variable fitness function provides a simple scheme for encoding a piecewise linear function into a genetic algorithm and a method for evolving these functions. The variable fitness functions are then used to determine the local fitness function at each step in the local search.…”

“…The Variable Fitness Function [1] describes how the weights of a weighted sum fitness function change over the iterations of a search process. The variable fitness function is piecewise linear, describing the relative importance of objectives at each iteration.…”

“…There are two variations: the standard variable fitness function fixes the number of discontinuities and the number of iterations between them and the adaptive variable fitness function allows the points of discontinuity to evolve along with the variable fitness function objective weights (Figure 1). Work so far provides evidence that the adaptive version is more effective than the fixed version, as more complex functions can be evolved [1].…”

“…The result is that the local fitness function is different from the global fitness function and can be more effective than the global fitness function to assess local moves. [1] shows the variable fitness function's effectiveness at enhancing local search heuristics and in this paper we attempt to show its ability to enhance a metaheuristics and to learn reusable information to guide the search of a difficult optimization problem. The problem we study is a complex real world workforce scheduling problem which contains many scheduling problems from the literature as subproblems.…”