Otimização multiobjetivo e programação genética para descoberta de conhecimento em engenharia

Abstract

The area of optimization involves the study and the use of methods to determine the parameters that lead to optimal solutions, according to criteria called objectives. A problem is classified as multiobjective when it presents multiple and conflicting objectives which must be simultaneously optimized. Recently, the interest of the researchers has grown in the analysis of post-optimality, which consists in the search for intrinsic properties of the optimal solutions of optimization problems. This can shed a new light on the understanding of the optimization problems. Innovization (from innovation through optimization) is a process of knowledge discovery from optimization problems in the form of mathematical relationships between variables, objectives, constraints, and parameters. Genetic Programming (GP), a search technique that can be used in this process, is a bio-inspired metaheuristic capable of evolving programs automatically. In addition to be numerically valid, the models found must correctly use the decision variables with respect to the units involved, in order to present coherent physical meaning. In this work, a method is proposed to handle the units through protected operations which ignore invalid terms. Also, a strategy is proposed here to avoid trivial solutions that do not add knowledge about the problem. In order to increase the diversity of the models obtained, it is also proposed the use of an external file to store the solutions of interest found during the search. Computational experiments are presented using five case studies in engineering to verify the influence of the proposed ideas. The problems dealt with here are the designs of: a 2-bar truss, a welded beam, the cutting of a metal part, composite gears, and a 10-bar truss. The latter was not previously explored in the knowledge discovery literature. Finally, the inferred knowledge in the case study of the 10-bar truss structure is used to reduce the dimensionality of that problem.