Chimera Objective #455
Chimera Objective #455
Conversation
|
|
|
@AVHopp @AdrianSosic PR in draft mode, NOT for review |
|
|
||
| # Force that every drawn target has a linear transformation and mode MIN or MAX. | ||
| assume(all(t.mode is not TargetMode.MATCH for t in targets)) | ||
| n_targets = len(targets) |
There was a problem hiding this comment.
not really sued often hence this shortcut is not needed
| # Normalize external merits to match internal implementation | ||
| ext_range = np.amax(ext_merits) - np.amin(ext_merits) | ||
| if ext_range > 0: | ||
| ext_merits = (ext_merits - np.amin(ext_merits)) / ext_range | ||
| else: | ||
| ext_merits = np.zeros_like(ext_merits) # Handle uniform values |
There was a problem hiding this comment.
@Scienfitz The reference code performs final normalization in scalarize via:
if np.amax(merits) > 0.:
merits = (merits - np.amin(merits)) / (np.amax(merits) - np.amin(merits))
See reference: reference line 259 onwards
This approach has two issues:
- It is not numerically stable: Division by (max - min) can cause division-by-zero when values are nearly identical
- When max(merits) ≤ 0, no normalization occurs, leaving merits in arbitrary negative ranges, causing confusions as they are acquisition weights for BoTorch (in minimization mode), our implementation always normalizes to [0,1] when range > 0 and handling edge case of a uniform merits
There was a problem hiding this comment.
@Scienfitz this however does not mitigate the warning message when running the test:
tests/test_objective.py::test_chimera_merits
/Users/x272626/Desktop/chimera/src/chimera/chimera.py:263: RuntimeWarning: invalid value encountered in divide
merits = (merits - np.amin(merits)) / (np.amax(merits) - np.amin(merits))
which means we have uniform merits (all values identical) -> this is however not sth that we can explicitly predict based on the input values, nor build the exclusion logics explicitly in the test as many different combinations and variations in target_vals and threshold_vals and threshold_type can lead to this issue.
How can we overcome this?
|
|
||
| target_vals = data.draw( | ||
| data_frames(columns=columns, index=range_indexes(min_size=2)).filter( | ||
| lambda df: len(df.drop_duplicates()) > 1 |
There was a problem hiding this comment.
Is this necessary? Because it could have multiple measurements that are the same
3 participants
This PR introduces Chimera, a general-purpose achievement scalarizing function for multi-target optimization that allows users to establish a hierarchy of targets with relative or absolute thresholds for concurrent optimization.
This implementation includes a new objective class,
ChimeraObjective, which follows a similar approach to theDesirabilityObjective. It scalarizes multiple targets into a single score, termedChimera Merits, which is to be minimized.For further details, please refer to the following publication:
F. Häse, L.M. Roch, and A. Aspuru-Guzik. Chimera: enabling hierarchy-based multi-objective optimization for self-driving laboratories. Chemical Science 2018, 9(39), 7642-7655.
WIP:
ChimeraObjectiveimplementation.DesirabilityObjective.ChimeraObjective.