Abstract:

To improve the accuracy and reliability of the drivability evaluation method,this study proposed a subjective and objective comprehensive evaluation method integrating the Extreme Gradient Boosting (XGBoost) algorithm, the Sparrow Search Algorithm (SSA) and the Shapley Additive Explanations (SHAP). Focusing on vehicle starting conditions, the study defined nine objective evaluation indicators to refine the drivability assessment system for vehicle start-up performance. A bidirectional mapping model between objective metrics and subjective scores was established using the XGBoost algorithm. To avoid local optima in the drivability evaluation model, the SSA was employed to efficiently optimize the core hyperparameters of XGBoost, thereby enabling the model to iteratively self-improve as the dataset expands. Finally, the SHAP was used to attribute the features of the mapping model, quantify the influence weight of objective evaluation indicators on the evaluation of drivability, and construct a comprehensive evaluation model of drivability with prediction accuracy, stability and interpretability. The proposed method was validated through multiple road tests that incorporate both domestic and international mainstream drivability evaluation frameworks. Comparative analysis shows that the proposed model outperforms mainstream approaches such as BP neural networks, random forests, and extreme learning machines (ELMs) in terms of MAE （Mean Absolute Error）, RMSE (Root Mean Square Error), and the coefficient of determination. The mapping accuracy is significantly improved, and the method’s interpretability makes it a valuable reference for integrating subjective and objective assessments in drivability evaluation.

Key words: drivability, starting condition, evaluation method, limit gradient lifting algorithm, Shapley additive explanations