Monte Carlo acquisition functions #48
Conversation
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
|
The |
| def probabilities(self, x, n_samples=256): | ||
| def probabilities(self, x, n_samples=1024): |
There was a problem hiding this comment.
What caused the default to change?
There was a problem hiding this comment.
Oops, this should probably be lower. It samples a Dirichlet distribution n_samples times in constructing the classification model.
| monte_carlo_upper_confidence_bound: | ||
| default_args: | ||
| beta: 4 | ||
| description: The expected value, plus some multiple of the uncertainty (typically \mu + 2\sigma). |
There was a problem hiding this comment.
Describe the MC effect/benefit here.
There was a problem hiding this comment.
Should I explain it here? It applies to all the MC acquisition functions (which is that it run faster, more robust and can run in parallel), which we should definitely put somewhere.
| acq_func_identifier: which acquisition function to use | ||
| n: how many points to get |
There was a problem hiding this comment.
Let's make it numpydoc-compatible: https://numpydoc.readthedocs.io/en/latest/format.html#parameters.
There was a problem hiding this comment.
I'll put numpydoc docstrings on all the agent methods in the next PR
Monte Carlo acquisition functions
This PR adds more Monte Carlo acquisition functions, which should be preferred to analytic ones (as they are faster to optimize and more robust).