AN EFFICIENT FRAMEWORK FOR OPTIMIZATION OF NONLINEAR STEEL TRUSSES WITH CONTINUOUS VARIABLES USING LIGHTGBM
Keywords:
Optimization, Truss, Direct Analysis, Nonlinear, Machine learningAbstract
Excessive time-consuming has been a great problem when applying metaheuristic algorithms to the optimization of structures using direct analyses, including steel truss structures. In this work, a robust optimization method was proposed to solve this issue. In the proposed method, an efficient variant of the differential evolution algorithm (DE), which was proved to be powerful in searching optimum designs and converged quickly, was employed as the optimizer. An effective framework based on LightGBM classification models was developed to save lots of time-consuming direct analyses required for evaluating constraints. To enhance the performance of LightGBM models, an adaptive parameter, which can reflect the convergence speed of the population, was proposed to prevent the imbalanced classification problem in the training data. Two truss optimizations with continuous design variables were studied, including a 47-bar power line and a 113-bar planar bridge. The results proved that the proposed method yielded better optimum designs than DE/best/1 and EpDE. It also saved more than 60% of time-computing compared to DE/best/1, EpDE, and 2EpDE. Besides that, the proposed framework for using LightGBM models was compatible with metaheuristics having different convergence rates, including DE/best/1, EpDE, and 2EpDE.