Cyber-physical systems (CPS) often rely on learned surrogate models whose performance degrades under non-stationary conditions. While many adaptive classifiers exist, adaptive methods for tabular regression are scarce. We propose UCAT, an adaptive model tree for online regression on non-stationary data streams in CPS. UCAT replaces the Hoeffding bound used by existing adaptive trees with a one-sided Wilcoxon–Mann–Whitney U-test on absolute prediction errors to select splits, and combines residual model trees with local linear models, Page-Hinkley drift detection, rival branches for subtree replacement, and exposure-based pruning. An optional twig-level threshold adaptation further refines split thresholds. On the SiD2Re benchmark (15 datasets, 7. 680 stream variants), we compare UCAT and a twig-adapting variant (UCATₜwig) with FIMT-DD using prequential evaluation. UCAT achieves lower cumulative absolute error on 68% of streams, with an average reduction of 13. 33%, and is preferred by normalized error metrics on most datasets, whereas FIMT-DD performs best mainly in stationary or high-noise regimes. These gains come at increased computational cost: UCAT is about 43% slower than FIMT-DD. Overall, UCAT provides an interpretable, streaming-capable regression method with statistically grounded split selection and targeted adaptation for CPS, suitable when improved accuracy and drift handling justify higher runtime.
Stratmann et al. (Thu,) studied this question.