Speciation, Clustering and other Genetic Algorithm Improvements for Structural Topology Optimization


Genetic algorithms are used to search for optimal structural topologies. Modifications to basic genetic algorithm techniques are implemented to increase computational efficiency, avoid premature convergence to a single solution, and solve new categories of problems. GA's are a search and optimization tool based on the principles of evolution and survival of the fittest. Potential designs are represented by chromosomes, each of which receives a fitness score based on the quality of the design it represents. As in nature, the more fit a chromosome is, the more likely it is to survive and "reproduce," combining with another "parent" chromosome to produce new "child" chromosomes. As this process of evaluation, selection, and reproduction is iterated, the population of chromosomes "evolves," and new and improved designs are generated. In this study, an allowable design domain is discretized into several binary, material/void elements, yielding a combinatorial search space. One chromosome represents one design, which can be evaluated using finite element analysis or analytical techniques. Extending previous efforts with the same basic representation and search technique, this research proposes several methods for improving genetic algorithm performance. New population initialization and parent selection methods are implemented to reduce run-times and decrease the number of poor intermediate designs generated and evaluated by the algorithm. Fitness sharing and speciation are used to distribute subsets of the evolving population over many local optima, preventing premature convergence to a single solution when multiple, equally good solutions exist. The resulting distribution of sub-populations is analogous to different species exploiting different niches in an ecosystem. Statistical cluster analysis techniques are used to divide the population into sub-species and to quantify the extent to which a population is speciated. Additionally, this cluster analysis is used to discourage the mating of dissimilar designs (designs from different clusters). Results show that these modifications to basic genetic algorithm techniques result in shorter run-times and greater diversity of solutions. Finally, a hybrid GA / simulated annealing method is introduced for topology design of adaptive structures. Thesis Supervisor: Mark J. Jakiela Title: Associate Professor, Mechanical Engineering

Cite this paper

@inproceedings{Duda2007SpeciationCA, title={Speciation, Clustering and other Genetic Algorithm Improvements for Structural Topology Optimization}, author={James W. Duda and Mark J. Jakiela}, year={2007} }