Add feasibility builder for explicit constraints #660
Conversation
Includes two new classes for linear and nonlinear constraints.
Include constraints and ctol.
Also added example notebook showing penalty-EI method.
| @@ -468,6 +468,84 @@ def acquisition(x: TensorType) -> TensorType: | |||
| return acquisition | |||
|
|
|||
|
|
|||
| class FastConstraintsFeasibility(SingleModelAcquisitionBuilder[ProbabilisticModel]): | |||
| """ | |||
| Uses the :func:`fast_constraints_feasibility` function to build a | |||
There was a problem hiding this comment.
I don't think this needs to mention that it uses fast_constraints_feasibility (which after all is an implementation detail)?
| Uses the :func:`fast_constraints_feasibility` function to build a | |
| Builder for a probability of feasiblity acquisition function that uses the residuals of | |
| direct constraints defined in the search space. |
There was a problem hiding this comment.
I copied that style from ProbabilityOfFeasibility, but I can remove it.
There was a problem hiding this comment.
Done in 1e41a57.
| default is CDF of the Normal distribution with a small scale. | ||
| """ | ||
| self._search_space = search_space | ||
| self._smoothing_function = smoothing_function |
There was a problem hiding this comment.
Should we check that search_space.has_constraints here and raise an exception if it doesn't?
There was a problem hiding this comment.
Yes I'll add that in.
There was a problem hiding this comment.
Done in 1e41a57.
| :param search_space: The global search space over which the feasibility of the constraints | ||
| is defined. | ||
| :param smoothing_function: The smoothing function used for constraints residuals. The | ||
| default is CDF of the Normal distribution with a small scale. |
There was a problem hiding this comment.
(explicitly state the default scale so people don't have to dig through the code?)
There was a problem hiding this comment.
Done in 1e41a57.
| @@ -15,7 +15,7 @@ | |||
| from __future__ import annotations | |||
There was a problem hiding this comment.
Do we also want to also add an integration test case to test_bayesian_optimization.py? (something similar to the notebook setup perhaps)
There was a problem hiding this comment.
This will take a bit longer to implement. I'll look into it and maybe add it as part of the next PR where we will have more optimisation methods.
| grid_density=30, | ||
| contour=True, | ||
| ) | ||
| plot_bo_points(query_points, ax[0, 0], num_initial_points, arg_min_idx) |
There was a problem hiding this comment.
The plot is a bit unclear for me: maybe make the dots a bit bigger and the purple one brighter? (Alternatively, ask @vpicheny for a second opinion!)
There was a problem hiding this comment.
That is an existing plotting function. I don't think the point size can be changed from its args but the colour can be, which I now have. Also, I have made the overall figure larger. Hopefully that is good enough.
| :return: The function for probability of feasibility of constraints. | ||
| """ | ||
|
|
||
| @tf.function |
There was a problem hiding this comment.
Looking at the notebook, it looks like this is being retraced every loop, which is not ideal! I think the issue is because ExpectedConstrainedImprovement calls the constraint function with the query points, which change shape each time. We may be able to specify an input_signature here to handle this. If you can't quickly figure it out don't worry though: I'll raise an issue since I expect it's happening for the normal ProbabilityOfFeasibility too.
There was a problem hiding this comment.
Yes, ProbabilityOfFeasibility has that problem too. I did spend a little bit of time looking at it, but didn't try changing the signature.
There was a problem hiding this comment.
I've raised an issue, thanks: #661
Improve documentation text Raise if search_space has no constraints Improve plot of BO points in notebook
Also includes initial notebook showing penalty-EI method. This notebook will be expanded further in the PR for constrained optimisation.