Bayesian Additive Regression Trees(BART) are a powerful Bayesian nonparametric regression model known for their high predictive accuracy and uncertainty quantification capabilities. However, they possess a limitation in achieving sufficient smoothing ...
Bayesian Additive Regression Trees(BART) are a powerful Bayesian nonparametric regression model known for their high predictive accuracy and uncertainty quantification capabilities. However, they possess a limitation in achieving sufficient smoothing performance due to their tree structure, which approximates functions using piecewise constant values. To address this limitation, various extension models have been proposed, utilizing techniques such as integration with Gaussian Process or the implementation of soft splitting structures. This paper aims to evaluate these extension models across diverse simulation scenarios to determine which models offer superior performance for nonlinear and smooth function estimation. The simulation results demonstrate that SoftBART exhibits stable predictive power in terms of accuracy, whereas GP-BART demonstrates superior performance in uncertainty-based metrics, specifically the Continuous Ranked Probability Score(CRPS). Finally, this paper evaluates the practical performance of these models using real data and examines how these extensions mitigate the structural limitations of standard BART and propose criteria for model selection tailored to specific analytical objectives.