Structural Risk Optimization Methods Based On Posterior Preference

Design optimization under uncertainty is aimed at guaranteeing the life-cycle performance of the product. The undesirable result of uncertainty is risk, which is inevitable for the realized design, hence risk optimization is tailor-made for engineering practice. This dissertation makes attempts to develop sophisticated risk optimization methods from a novel optimization and decision-making perspective. Considering the nested nature of risk optimization, this dissertation focuses on the following three key points:1) Based on generalized theory of uncertainty, this dissertation extends design optimization under uncertainty to risk-based cases with granular probability and posterior preference. Since risk optimization with posterior preference is inherently a many-objective problem, uncertainty management is realized via adjusting the number of granular probabilities and refining the uncertainty modeling, while uncertainty mitigation can be realized under fixed number of objectives. To facilitate optimization and decision-making, multi-objective evolutionary algorithm based on decomposition (MOEA/D) in conjunction with a filtering criterion is employed. The competence of MOEA/D lies in its capability on reproducing the total cost after optimization for multi-attribute decision making, hence it is compatible with structural risk optimization with posterior preference.2) The convergence speeds of multi-objective evolutionary algorithms for solving truss multi-objective topology optimization (MOTO) are studied rigorously and empirically. The Pareto fronts for widely collected and modified MOTO benchmarks are rigorously derived using enumeration. Then this dissertation proposes a general performance assessment method tailor-made for examining the convergence and efficiency of MOEAs on solving MOTO. In the subsequent comparative study, the limit of MOEAs’convergence speed on solving MOTO is revealed by comparing eight state-of-the-art MOEAs. Based on the multi-level convergence requirement, the convergence and efficiency can be evaluated in an integrated manner. Besides, this dissertation reveals some unveiled difficulties posed by truss MOTO problems. In this way, this dissertation provides sound algorithmic bases for risk optimization with posterior preference in the deterministic optimization loop.3) Considering there is temporarily no available model for the loss estimation of wind-induced hazard, this dissertation proposes a risk optimization method with posterior preference for wind-resistant design of tall buildings. Closed-form solution for3D wind loads is employed for uncertainty quantification and propagation, and kernel-learning based principle component analysis is employed for objective dimensionality reduction. A hybrid micro multi-objective particle swarm optimization is proposed to solve the many-objective optimization problem. This dissertation empirically proves that the bi-objective problem involving initial cost and the failure probability of most unfavorable RMS acceleration response approximates the original many-objective problem well. This way, the original problem is simplified from a complex multi-objective decision making process into a relatively easy multi-attribute decision making process. Accordingly, convincing decisions can be made based on the explicit building performance rather than the unreliable loss information.To conclude, this dissertation preliminarily establishes risk optimization methods based on posterior preference from theory, algorithm and its application to wind-resistant design.