Evolutionary and genetic strategies for topology optimization of frameworks : pareto-comparisons and hybrid methods

The main objective of this research is to develop new efficient and cost-effective topology optimization methods for framework structures. The first part of the thesis concentrates on the assessment of the effectiveness of the Evolutionary Structural Optimization (ESO) method of designing frameworks. This method is critically examined by studying the trajectory in which designs evolve during the ESO process on the weight-maximum stress plane and later overlaying it with the Pareto Front (PF) for the two-objective problem of simultaneously minimising weight and the maximum stress within the structure. To study the Pareto-efficiency of the ESO method, the designs obtained by ESO are compared with the designs obtained using exhaustive search for combined performance on two counts: the maximum stress within the structure and the overall weight. Whilst for complex problems an exhaustive search is not practical, the approach adopted here is to encode the problem formulation using a genetic algorithm (GA) and to allow the formulation to evolve in the direction of improving Pareto optimal designs. Since GA is a stochastic method, the robustness of the conclusions has been assessed by running GA with multiple seeds. Numerical experiments show that ESO produces reasonable designs at little expense; however, the procedure misses out on several efficient designs if one could afford the computational expense. As far as topology and size optimization is concerned, it is observed that ESO produces Pareto sub-optimal designs, but is superior to GA if one could not afford a computationally demanding search. ESO is computationally efficient but it fails to produce some designs with very good structural performance. In the second part of the thesis two new strategies for topology optimization of frameworks are developed by combining the Evolutionary Structural Optimization (ESO) method and Genetic Algorithms (GA). This approach combines the quality of a stochastic global search such as GA and the computational efficiency of ESO. The first method proposed here is the ESO assisted GA method (ESOaGA) in which ESO obtained designs are inserted in the GA population, helping the GA search to operate in more promising directions. The second method is the GA assisted ESO method (GAaESO), in which GA produced designs are used as starting points for a family of ESO runs. The designs obtained by the proposed methods and the “unassisted” GA are compared visually and quantitatively using three quality indicators: the hypervolume, epsilon and R indicators. The statistical significance of the quality indicators is also assessed. Again, two goals are used in this comparison: the maximum stress within the structure and the overall weight. Both hybrid methods can obtain better optimized designs in less computational time than the respective “unassisted” methods. Finally, an iterative application of GA and ESO is explored.EThOS - Electronic Theses Online ServiceGBUnited Kingdo