Publications

A benchmark of optimization algorithms for thermal, luminous and energy multi- objective analysis on Grasshopper for Rhino. Advisor: Joyce Correna Carlo. Co- advisor: Rafael de Paula Garcia. This master thesis aims to establish which algorithm is more suited to a Simulation- based optimization (SBO) process, based on the type of simulation used, and also on the number of parameters, type of parameters, and number of fitness functions. We focused on the optimization algorithms for multi-objective optimization processes, since they have a fundamental role in SBO processes. We choose to use the Grasshopper for Rhinoceros platform due to its diversity and robustness, that allows performing parametric modelling, simulation, and optimization in the same environment. We initiate this study investigating the optimization engines available on Grasshopper and decide to focus on Opossum and Octopus. For multi-objective optimization, Opossum has RBFMOpt, NSGA2, MOEA/D, NSPSO, and MHACO, and Octopus has HypE and SPEA2. Then, we used seven different algorithms. We proposed 14 building performance related problems. The problems varied from 5 to 18 parameters, and required at least one type of simulation such as thermal, luminous, and energy. We compare the algorithms’ performance by using Python implementations of different performance metrics, such as hypervolume, modified inverted generational distance, generational distance, and additive epsilon indicator, that provided a robust methodology to assess algorithms’ performance and state which one is more suited for each optimization problem. We also applied the Kruskal-Wallis non-parametric test to support stating the difference between algorithms performance and also to assess the potential of each algorithm to computational cost reduction. Based on this benchmark steps, we initially compared the performance of RBFMOpt, NSGA2, and MHACO on a single problem. Then we advance by proposing a sequential study with all algorithms and nine problems. The overall results point out that RBFMOpt has the best performance, especially with its default hyperparameters configurations. RBFMOpt not only provides the best results but also need less function evaluations to obtain those results, and also presents an additional tendency for computational cost reduction by allowing reducing the number of runs withoutsignificantly impact its average performance. HypE also have a good performance, with the second position on the overall ranking, but requires more function evaluations than RBFMOpt. In general, RBFMOpt should be used in multi-objective SBO processes in the Grasshopper platform, especially when the simulator has a lower budget or more time cost consuming simulations.

Keywords: Benchmarking. Simulation-based optimization. Model-based algorithm. Bioinspired algorithms. Performance metrics. Building performance simulation.

Dissertação: clique aqui