A Data-Oriented Optimization Framework

Abstract

Design optimization using the Multi-Objective Evolutionary Algorithm (MOEA) has still been studied, progressed well, and used to improve building performance. Besides, floor plan generation that is the problem of fitting several rooms into an outline given beforehand has recently been studied well using machine learning models. Although the building performance and a floor plan intimately relate, they are rarely combined in one optimization framework. A separation of these problems often forces users to manually explore accurate floor plans in a solution space or limit optimizing the building performance following certain machine learning methods and its dataset. We mainly focused on these issues and developed a custom-made model that contains association rule mining and the cosine similarity formula extracted from machine learning methods. This model of lazy learning is added to an MOEA-based optimization framework and outputs the total cosine similarity between each generated floor plan and the referred plans dynamically selected from our dataset, and the framework maximizes it. We applied this framework to a case study on generating eco-conscious office building designs that will enable them to convert easily in the future. This paper elaborates on how to create a dataset and formulation for optimization, and we emphasize the plausibility of floor plan generation. Finally, we demonstrated the efficiency of the framework by comparing the performance indicators of optimization.