The California bearing ratio (CBR) is one of the important indexes, which is used to represent the strength of subgrade or subbases of pavement. In general, the CBR can be determined through experiments both in the laboratory and field. However, the determination of the CBR is time and cost-consuming as well as low accuracy due to the disturbance of samples and limitations of preparation in the laboratory. Thus, this study estimates the CBR of stabilized soil using twelve machine learning techniques (6 single models and 6 hybrid models). The single models include artificial neural network (ANN), gradient boosting (GB), extreme gradient boosting (XGB), random forest (RF), support vector machine (SVM), and K-nearest neighbors (KNN), while the six hybrid models are a combination of these single models and random restart hill-climbing optimization (RRHC). Twelve models are constructed based on eleven input variables, including cement, Atterberg's limits, optimum moisture content (OMC), maximum dry density (MDD), and dust and ashes. To evaluate the performance of the proposed models, four popular statistical indexes namely mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and determination coefficient (R²) were used. The results indicate that among twelve models, four models using GB, RRHC_XGB, RRHC_RF, and RF had a high prediction accuracy (R² > 0.98) and outperformed than other remaining models. Among these four models, the model using RF has the highest prediction accuracy (R² = 0.9817, RMSE = 2.3970, MAE = 1.1682, MAPE = 0.0666). According to the result of feature importance analysis using Sklearn permutation importance, SHapley Additive exPlanation (SHAP), individual conditional expectation (ICE), and partial dependence plots-2D, cement and plasticity index (PI) are two most important variables affecting the CBR prediction of stabilized soil. Where, PI was found to be a very crucial factor, which improved the prediction ability of the models compared to the results of previous studies. The results also reveal that when cement content is larger than 2%, there is an insignificant influence of the cement on the CBR of stabilized soil. The value of PI smaller than 15% has a vital impact on the CBR for any values of cement, dust, and ashes. Furthermore, the results also indicate that dust and ash content have less effect on the CBR of stabilized soil. In summary, it can be concluded that this study provides an insightful assessment of the CBR prediction of stabilized soil, and the results of this study can fill the gap in the literature and provide practical knowledge and application on the CBR of stabilized soil.

加州承载比（CBR）是重要指标之一，用来表示路面的路基或底基层的强度。一般来说，CBR 可以通过实验室和现场的实验来确定。然而，由于样品的干扰和实验室制备的限制，CBR的测定既费时又费钱，而且准确性低。因此，本研究使用 12 种机器学习技术（6 个单一模型和 6 个混合模型）估计了稳定土的 CBR。单个模型包括人工神经网络 (ANN)、梯度提升 (GB)、极端梯度提升 (XGB)、随机森林 (RF)、支持向量机 (SVM) 和 K-最近邻 (KNN)，而六种混合模型是这些单一模型和随机重启爬山优化（RRHC）的组合。12 个模型基于 11 个输入变量构建，包括水泥、Atterberg 极限、最佳含水量 (OMC)、最大干密度 (MDD) 以及灰尘和灰烬。为了评估所提出模型的性能，四个流行的统计指标分别是平均绝对误差 (MAE)、均方根误差 (RMSE)、平均绝对百分比误差 (MAPE) 和决定系数 (R² ) 被使用。结果表明，在 12 个模型中，使用 GB、RRHC_XGB、RRHC_RF 和 RF 的 4 个模型具有较高的预测精度（R ²  > 0.98）并且优于其他剩余模型。在这四个模型中，使用 RF 的模型具有最高的预测精度（R ² = 0.9817，RMSE = 2.3970，MAE = 1.1682，MAPE = 0.0666）。根据使用 Sklearn 置换重要性进行特征重要性分析的结果，SHApley Additive exPlanation (SHAP)、个体条件期望 (ICE) 和 2D 部分依赖图，水泥和塑性指数 (PI) 是影响 CBR 预测的两个最重要的变量的稳定土。其中，PI 被发现是一个非常关键的因素，与以前的研究结果相比，它提高了模型的预测能力。结果还表明，当水泥含量大于2%时，水泥对稳定土CBR的影响不显着。PI 值小于 15% 对水泥、灰尘和灰烬的任何值的 CBR 都有至关重要的影响。此外，结果还表明，粉尘和灰分含量对稳定土的CBR影响较小。综上所述，可以得出结论，本研究对稳定土的 CBR 预测提供了深刻的评估，本研究结果可以填补文献空白，为稳定土的 CBR 提供实用知识和应用。