Restart Strategies for Constraint-Handling in Generative Design Systems

Abstract

Product alternatives suggested by a generative design system often need to be evaluated on qualitative criteria. This evaluation necessitates that several feasible solutions which fulfill all technical constraints can be proposed to the user of the system. Also, as concept development is an iterative process, it is important that these solutions are generated quickly; i.e., the system must have a low convergence time. A problem, however, is that stochastic constraint-handling techniques can have highly unpredictable convergence times, spanning several orders of magnitude, and might sometimes not converge at all. A possible solution to avoid the lengthy runs is to restart the search after a certain time, with the hope that a new starting point will lead to a lower overall convergence time, but selecting an optimal restart-time is not trivial. In this paper, two strategies are investigated for such selection, and their performance is evaluated on two constraint-handling techniques for a product design problem. The results show that both restart strategies can greatly reduce the overall convergence time. Moreover, it is shown that one of the restart strategies can be applied to a wide range of constraint-handling techniques and problems, without requiring any fine-tuning of problem-specific parameters.