Many engineering problems require both domain knowledge and search/optimization techniques for their solution. Hybrid systems that use robust search algorithms when confronted with a problem outside the scope of their knowledge, bring up the issue of how best to incorporate this knowledge in guiding robust search. This paper studied the problem of incorporating domain knowledge in genetic algorithm search for the configuration design and optimization of trusses.
We used domain knowledge to initialize the genetic algorithm's population and to guide the application of crossover and mutation. This substantially reduces the search space and speeds convergence. The genetic algorithm is able to produce feasible, useful solutions in approximately twenty minutes on a high-end workstation. Broadly speaking, the GA first settles on a topology and geometry, and spends the rest of the time optimizing member cross sections.
This paper explores an essential aspect of interfacing genetic search with knowledge-bases. As a next step, we plan to use case-based reasoning to store genetic algorithm solutions and use these stored cases to initialize the population.
The representation used in this paper allows easy incorporation of design heuristics and constraints. This leads to a smaller search space and therefore fast convergence times. However, relaxing the constraints and expanding the search space may lead to more exploration and perhaps better and more innovative solutions [Louis, 1993]. We are currently exploring this tradeoff. One of the problems in our representation is that crossover can generate structures that are not viable; in other words, structures that cannot bear the load or that are simply not realistic. This causes the finite element analysis program to abort. Our current solution is to patch up such structures by adding bracing members. A representation that guarantees viable offspring would be better. Shape grammars [Stiny and Gips, 1978] offer an alternative representation and have been used by Cagan and Mitchell.
In the future, we plan to investigate the influence of various program parameters on the results, which include the penalty weights for overstressing and understressing, the probability of crossover and mutation, and parameters reflecting design heuristics and constraints. We also plan to develop different representations which are more easily manipulated by the genetic algorithm as well as being more efficient. Along with representations, we can try different search algorithms such as simulated annealing or other stochastic hill-climbers and explore the encoding of domain knowledge for these algorithms. Finally, we hope to extend the methodology to other types of structural system design as well as system configuration design in different domains.