Doctor of Philosophy
Other advisers/committee members
Chen, Brian Y.; Terlaky, Tamas; Wild, Stefan M.
Optimization problems with numerical noise arise from the growing use of computer simulation of complex systems. This thesis concerns the development, analysis and applications of randomized derivative-free optimization (DFO) algorithms for noisy functions. The first contribution is the introduction of DFO-VASP, an algorithm for solving the problem of finding the optimal volumetric alignment of protein structures. Our method compensates for noisy, variable-time volume evaluations and warm-starts the search for globally optimal superposition. These techniques enable DFO-VASP to generate practical and accurate superpositions in a timely manner. The second algorithm, STARS, is aimed at solving general noisy optimization problems and employs a random search framework while dynamically adjusting the smoothing step-size using noise information. rate analysis of this algorithm is provided in both additive and multiplicative noise settings. STARS outperforms randomized zero-order methods in both additive and multiplicative settings and has an advantage of being insensitive to the level noise in terms of number of function evaluations and final objective value. The third contribution is a trust-region model-based algorithm STORM, that relies on constructing random models and estimates that are sufficiently accurate with high probability. This algorithm is shown to converge with probability one. Numerical experiments show that STORM outperforms other stochastic DFO methods in solving noisy functions.
Chen, Ruobing, "Stochastic Derivative-Free Optimization of Noisy Functions" (2015). Theses and Dissertations. 2548.