| Summary: | This dissertation is framed in the research project "BoostDFO: Improving the performance and moving to newer directions in Derivative-Free Optimization", funded by Fundação para a Ciência e Tecnologia, whose objective is to develop efficient and robust algorithms for Global and Multiobjective Derivative-Free Optimization. The demand for algorithms belonging to this class appears in different application areas, such as robotics, electrical engineering, aeronautics or oceanography, where, for several reasons, it is not possible to use derivatives. Global and Local Optimization using Direct Search (GLODS) is an algorithm belonging to this class of optimization methods, which aims at identifying the global minimum of a given problem by computing all the local minima. However, identifying points as global minimizers is always a hard task, being of increased complexity when the function to optimize is computational expensive and time-consuming. Typically, the higher the dimension and complexity of the problem, the more computational effort and time will be required to run the algorithm. The current version of GLODS is developed sequentially, and it is not fully optimized. The present work will analyze in detail the execution and behavior of GLODS, and will propose and evaluate the numerical performance of different parallelization strategies, implemented using the MATLAB Parallel Computing Toolbox (PCT). The parallelization structure will have a primary purpose of allowing to distribute objective function evaluations among different processors of the hardware platforms, both host locally or in public clouds. Three strategies for GLODS parallelization were designed and implemented centered on the poll phase of the algorithm. The first two - parallelization of the poll step with one and two poll centers - were successful and gave good results in terms of solution quality and execution time reduction. The third level of parallelization introduced the possibility of dynamically selecting the number of poll centers, according with the number of processors available, a functionality that can be particularly useful in future work in the area.
|
|---|