Injecting problem-dependent knowledge to improve evolutionary optimization search ability

Abstract

The flexibility introduced by evolutionary algorithms (EAs) has allowed the use of virtually arbitrary objective functions and constraints—even when evaluations require, as for real-world problems, running complex mathematical and/or procedural simulations of the systems under analysis. Even so, EAs are not a panacea. Traditionally, the solution search process has been totally oblivious of the specific problem being solved, and optimization processes have been applied regardless of the size, complexity, and domain of the problem. In this paper, we justify our claim that far-reaching benefits may be obtained from more directly influencing how searches are performed. We propose using data mining techniques as a step for dynamically generating knowledge that can be used to improve the efficiency of solution search processes. In this paper, we use Kohonen SOMs and show an application for a well-known benchmark problem in the water distribution system design literature. The result crystallizes the conceptual rules for the EA to apply at certain stages of the evolution, which reduces the search space and accelerates convergence.