Transfer Learning

[silviu20]

Project Documentation

Project A3

Multi-objective optimization using Pareto optimization.

Optimization Configuration

Parameters

Name	Type	Values/Range	Options
P1	Numerical Continuous	0 to 200	-
P2	Numerical Continuous	0 to 100	-
P3	Numerical Discrete	10, 20, 30, 40, 50	-
P4	Numerical Discrete	90, 105, 120	-
P5	Categorical	Red, Blue, Green	Encoding: One-Hot (OHE)
P6	Categorical	Batch, Semi-Batch, Continuous	Encoding: One-Hot (OHE)

Targets

Name	Mode	Bounds
Target 1	MAX	0 to 100
Target 2	MAX	0 to 100

Constraints

No constraints configured.

Acquisition Function

qNoisyExpectedHypervolumeImprovement

Project B3_TL_A3

Transfer learning project using data from Project A3.

Issue

Issue occurs when seeking new experimental suggestions on Project B3_TL_A3:

Error Log:

ERROR - Error suggesting next point: _partition_space does not support batch dimensions.

raceback (most recent call last):
File "/app/src/core/optimization/backend.py", line 2872, in suggest_next_point
suggestions = campaign.recommend(
File "/usr/local/lib/python3.10/dist-packages/baybe/campaign.py", line 526, in recommend
rec = self.recommender.recommend(
File "/usr/local/lib/python3.10/dist-packages/baybe/recommenders/pure/bayesian/base.py", line 198, in recommend
self._setup_botorch_acqf(
File "/usr/local/lib/python3.10/dist-packages/baybe/recommenders/pure/bayesian/base.py", line 124, in _setup_botorch_acqf
self._botorch_acqf = acqf.to_botorch(
File "/usr/local/lib/python3.10/dist-packages/baybe/acquisition/base.py", line 84, in to_botorch
).build()
File "/usr/local/lib/python3.10/dist-packages/baybe/acquisition/_builder.py", line 172, in build
botorch_acqf = self._botorch_acqf_cls(**self._args.collect())
File "/usr/local/lib/python3.10/dist-packages/botorch/acquisition/multi_objective/monte_carlo.py", line 333, in init
NoisyExpectedHypervolumeMixin.init(
File "/usr/local/lib/python3.10/dist-packages/botorch/utils/multi_objective/hypervolume.py", line 645, in init
self._set_cell_bounds(num_new_points=X_baseline.shape[0])
File "/usr/local/lib/python3.10/dist-packages/botorch/utils/multi_objective/hypervolume.py", line 769, in _set_cell_bounds
self.partitioning = self.p_class(
File "/usr/local/lib/python3.10/dist-packages/botorch/utils/multi_objective/box_decompositions/non_dominated.py", line 382, in init
super().init(ref_point=ref_point, Y=Y)
File "/usr/local/lib/python3.10/dist-packages/botorch/utils/multi_objective/box_decompositions/box_decomposition.py", line 268, in init
super().init(ref_point=ref_point, Y=Y, sort=ref_point.shape[-1] == 2)
File "/usr/local/lib/python3.10/dist-packages/botorch/utils/multi_objective/box_decompositions/box_decomposition.py", line 63, in init
self.partition_space()
File "/usr/local/lib/python3.10/dist-packages/botorch/utils/multi_objective/box_decompositions/box_decomposition.py", line 319, in partition_space
super().partition_space()
File "/usr/local/lib/python3.10/dist-packages/botorch/utils/multi_objective/box_decompositions/box_decomposition.py", line 147, in partition_space
self._partition_space()
File "/usr/local/lib/python3.10/dist-packages/botorch/utils/multi_objective/box_decompositions/box_decomposition.py", line 330, in _partition_space
raise NotImplementedError(
NotImplementedError: _partition_space does not support batch dimensions.

The issue I see is that using TL within a Pareto type optimization does raise some issue regarding the acquisition function. For Project B3_TL_A3 I have to switch to another acquisition function (qLogNParEGO).

Is this just me or is this the intended design?

Thanks

Silviu

[Scienfitz]

potential bug please continue in #713