A common problem in computer-aided design (CAD) is that of constraint-based geometric design. Typically, one wishes to place a number of points in space subject to a given set of constraints, where the number of constraints is sufficient to determine a finite number of solutions. In general, this finite number of solutions is still exponential in the number of points; among these, the designer has one intended solution in mind. Genetic algorithms (GAs) provide a promising approach to finding this intended solution.

The genetic algorithm will have a number of parameters, including mutation probability, crossover rate, and population size, that need to be chosen so that the GA converges to the intended solution. This paper describes a system that learns a Bayesian network that can be used to suggest the required parameter set, and moreover chooses it in a problem-independent way. The results are compared to a detailed study performed by Barreiro, Joan-Arinyo, and Luzon [1]. The system described in this paper is able to learn the optimal solution as given by the statistical study. The authors also consider the changes in the Bayesian network, as more cases are added to the database. These kinds of studies are of considerable interest, and help to lay the groundwork for future work in the design of genetic algorithms.