Skip to main content

Heuristic models (train, test, run)

Heuristic models let you train once on simulated data and then predict asterism quality in milliseconds.

  • Mono-NGS: smooth splines fitted on distance/flux (with frequency bins).
  • Multi-NGS: a small neural network.

The model predicts a jitter/penalty score and derives Strehl, FWHM, and EE. Lower penalty = better asterism.

Quick start​

⚙️​ Train​

from tiptop import generateHeuristicModel
simul = generateHeuristicModel(simulName, path2param,        
                               parametersFile,
                               outputDir, outputFile, doPlot=False, doTest=True)

What this does:

  • Runs asterism evaluation if needed (or reloads it if already computed).
  • Trains a model:
    • mono: 3 SmoothBivariateSpline (frequency buckets) + static terms
    • multi: PyTorch MLP
  • Saves the model to disk in outputDir:
    • Mono: <parametersFile without .ini>_hmodel.npy
    • Multi: <parametersFile without .ini>_hmodel.pth

🔎 Test​

If doTest=True, a hold-out split is used. TipTop prints:

  • mean/median absolute and relative errors,
  • RMS error,
  • how often the model’s ranking matches the AO ranking (where separations are significant).

You can also call the low-level tester yourself (see below).

▶️ Run​

Once trained, you can call the model to rank new asterisms instantly, without full AO simulations.

results = simul.runHeuristicModel()
# `results` is a list of AsterismProperties, already sorted by predicted jitter asc:
# [AsterismProperties(index=..., asterism=[...], jitter=..., strehl=..., fwhm=..., encircled_energy=...), ...]

Tip: If you want the model to auto-load on construction, put this in your .ini:

[ASTERISM_SELECTION]
heuristicModel = mycase_hmodel   ; basename (no extension); looked up in outputDir

For mono, TIPTOP loads outputDir/mycase_hmodel.npy. For multi, use the .pth equivalent.