Model Competition

Instead of training one model type and hoping it's the right choice, Run Competition trains any combination of Random Forest, XGBoost, and Neural Network competitors concurrently against the same input/output channels - including auto-tuning for the tree-based competitors, not just Neural Network. Every competitor is scored on a held-out validation split it never trained on, and the leaderboard shows validation RMSE and R² side by side for each entry, along with training time and a fit diagnosis (underfitting, overfitting, or genuinely competitive). For datasets large enough to afford it, a second, untouched test split - never used during training or model selection - gives a more trustworthy accuracy number for whichever entry you plan to ship. Any entry on the leaderboard can be exported, not just the winner - though exporting to code requires a Virtual Sensor license, while training and comparing competitors themselves only needs the Viewer tier.

While a competitor is still running, the number shown next to it means something different depending on its phase: the best hyperparameter combination found so far while tuning, or the winning configuration's own live, still-converging score once retraining begins. That number can briefly look worse right as retraining starts, before settling back past the tuning phase's best - expected behavior, not a step backward.