Fix bugs in tranquilo

[timmens]

• Make sure that centering and scaling is done via map_to_unit and map_from_unit
• Adjust model.py to the case of sphere and cube trustregions; focus on move_model.
• Add effective_center and effective_radius to Region
• Add residualize option
• Fix get_default_model_fitter
• Use state.x instead of state.trustregion.center when appropriate

Notes for reviewer:

• Check filter_points.py line 96-97:

  order = 2 if state.trustregion.shape == "sphere" else np.inf
  dists = np.linalg.norm(xs - state.x, axis=1, ord=order)

• Check new_history.py line 207-213:

  out = _find_indices_in_region(xs, center=region.center, radius=region.radius)
  
  if region.shape == "sphere":
      mask = np.linalg.norm(xs[out] - region.center, axis=1) <= region.radius
      out = out[mask]
  
  return out

[codecov]

Codecov Report

Merging #449 (6a14073) into main (3ac4dfb) will increase coverage by 0.25%.
The diff coverage is 98.98%.

❗ Current head 6a14073 differs from pull request most recent head 193523a. Consider uploading reports for the commit 193523a to get more accurate results

@@            Coverage Diff             @@
##             main     #449      +/-   ##
==========================================
+ Coverage   92.95%   93.21%   +0.25%     
==========================================
  Files         249      247       -2     
  Lines       18311    18018     -293     
==========================================
- Hits        17021    16795     -226     
+ Misses       1290     1223      -67     

Impacted Files	Coverage Δ
src/estimagic/optimization/pounders_history.py	91.66% <ø> (ø)
...timagic/optimization/tranquilo/aggregate_models.py	97.67% <ø> (ø)
src/estimagic/optimization/tranquilo/tranquilo.py	97.26% <ø> (ø)
tests/visualization/test_visualize_tranquilo.py	100.00% <ø> (ø)
src/estimagic/optimization/tranquilo/models.py	95.45% <93.33%> (+1.75%)	⬆️
src/estimagic/optimization/pounders.py	93.33% <100.00%> (ø)
...imagic/optimization/tranquilo/estimate_variance.py	100.00% <100.00%> (ø)
.../estimagic/optimization/tranquilo/filter_points.py	81.13% <100.00%> (+1.29%)	⬆️
src/estimagic/optimization/tranquilo/fit_models.py	86.27% <100.00%> (+1.75%)	⬆️
src/estimagic/optimization/tranquilo/history.py	94.59% <100.00%> (ø)
... and 19 more

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

[janosg]

About the two points you mention:

1. I think we can decide that history searches are always based on an l2 norm and thus get rid of line 96.

2. I think we always want to do what we currently do in the sphere case. We should check how much slower it is to just do the whole trustregion search with an l2 norm in numpy. (Background: We currently do a greedy search in numba that uses an infinity norm (i.e. no calculations, just if conditions) and stops as soon as the distance is large in any dimension).

[janosg]

Thanks a lot! Very nice.

A last general comment: We should probably replace the suffix _centered (e.g. for x, samples, ...) by _unit.

[janosg]

Very nice!