ABSTRACT

This study developed gradient decision tree boosting (GDTB) models to estimate dynamic moduli ($\left|E^{\ast }\right|$) of hot mix asphalt (HMA) mixtures. The GDTB used as input the binder properties, mixture volumetric, and aggregate gradation of the mixtures. The data used for training the GDTB were extracted from a report of the National Cooperative Highway Research Program (NCHRP) project 9-19 [Witczak, M., 2006. Simple performance tests: summary of recommended methods and database. Washington, D.C.: Transportation Research Board, No. 547 in NCHRP Report.]. Totally, 7400 records of data for 346 mixtures were involved, among which 6700 were randomly chosen for training, 200 for validation, and 500 for testing. Comparative analyses were conducted among the GDTB, the two Witczak's equations, and two neural networks (NNs). This study emphasized both the predictive accuracy and computation efficiency of the models. The results indicated that the GDTB achieved predictive accuracy that was significantly higher than the Witczak's models and was in parallel to the more complex NNs. Compared to the Witczak's equations, for the viscosity-based model, the GDTB increased the coefficients of determination ($R^2$) by 51.5% (arithmetic) and 11.5% (logarithmic), respectively; for the $\left|G^{\ast }\right|$ based model, it respectively increased the $R^2$ by 22.5% (arithmetic) and 8% (logarithmic). Besides the enhanced predictive accuracy, the GDTB only marginally increased the computing time comparing with the empirical equations.

摘要

本研究开发了梯度决策树增强 (GDTB) 模型来估计动态模量 ($\left|{乙}^{*}\right|$) 的热拌沥青 (HMA) 混合物。GDTB 用作混合物的粘合剂性质、混合物体积和骨料级配的输入。用于训练 GDTB 的数据是从国家合作公路研究计划 (NCHRP) 项目 9-19 [Witczak, M., 2006 年的报告中提取的。简单的性能测试：推荐方法和数据库的总结。华盛顿特区：交通研究委员会，NCHRP 报告中的第 547 号。]。共涉及 346 种混合物的 7400 条数据记录，其中随机选择 6700 条用于训练，200 条用于验证，500 条用于测试。在 GDTB、两个 Witczak 方程和两个神经网络 (NN) 之间进行了比较分析。本研究强调模型的预测准确性和计算效率。结果表明，GDTB 实现的预测精度明显高于 Witczak 模型，并且与更复杂的 NN 并行。与 Witczak 方程相比，对于基于粘度的模型，GDTB 增加了决定系数 ($R^2$) 分别降低 51.5%（算术）和 11.5%（对数）；为了$\left|G^{*}\right|$基于模型，它分别增加了$R^2$22.5%（算术）和 8%（对数）。除了增强的预测准确性外，与经验方程相比，GDTB 仅略微增加了计算时间。