Skip to content

Learning With FaMoS-DTW

Run from the repository root:

just examples-famos

This example learns sampled affine dynamics and next-location trees from the native thermostat, without simulator location labels. It uses unchanged thermostat physics and continuous event times, sampled on a uniform grid. All runs start in heating. Training and held-out initial temperatures differ; the target temperature is constant.

Edit the experiment settings in the source. It calls three independent stages:

  1. discover_famos_locations: geometric segmentation and exact per-signal DTW clustering, with partial labels for derivative warmup or insufficient evidence.
  2. fit_famos_dynamics: output-only affine ARX fits within discovered segments.
  3. fit_famos_transitions: next-location decision trees, one per source location, from explicitly labeled frames and ordered output/input features plus the declared location count. Use discovery.to_labeled_frames(training) to supply FaMoS assignments.

The tree fitter reads a nullable integer location_id column and current physical outputs followed by exogenous controls (which may be empty). It uses source-time features to predict the next row's location, including self transitions. Pairs with either label null are skipped without bridging gaps or independent traces; every declared source must have an eligible outgoing pair. No timestamps are inferred, and unrelated columns are ignored. Stored training-only maximum-absolute scales (zero replaced by one) normalize features; prediction accepts physical-unit values.

FamosModel composes the fitted artifacts for recursive prediction. Output/input feature order and location count must match, but training scales may differ because each artifact predicts in physical units. Matching location meanings and source-time training alignment remain the caller's responsibility; structural checks do not certify provenance. Explicit labels can also come from another learner, as in the fixed-dynamics MMSR comparison. Initialization uses one held-out temperature and a learned training location; subsequent measured temperatures are used only for scoring. Both the next output and next location use the current location, output and input. See the API reference for history shapes, scales and timing.

The script reports discovered locations, unassigned rows, fit ranks and training errors, then held-out recursive RMSE and maximum absolute error. It scores temperature predictions with NumPy after excluding the one supplied seed, and reports how many rows were scored. These are diagnostics, not accuracy acceptance tests. A constant target is collinear with the affine bias, so rank deficiency is expected. This experiment does not identify arbitrary target response: the native target changes switching thresholds, not flows. Segmentation errors, sampled approximations to continuous events, and tree extrapolation can all affect recursive predictions.

The method follows the FaMoS paper and official artifact. Optional LMI refinement is not included. Output-only ARX fitting and source-conditioned trees are explicit adaptations, not a MATLAB wrapper.