How should learners handle data that is outside of the domain

[basnijholt]

(original issue on GitLab)

opened by Jorn Hoofwijk (@Jorn) at 2018-09-19T09:25:27.165Z

Since giving a user the freedom to add points outside the domain would introduce some bugs that require careful rewriting of the code to mitigate.
Apart from that, IMHO it doesn't make much sense to add a point outside of the domain. Therefore I propose we raise an exception whenever a user tries to tell a point outside of the domain, because it then forces users to think about what they are doing if they do not use the provided adaptive runners.

[basnijholt]

originally posted by Jorn Hoofwijk (@Jorn) at 2018-09-19T14:13:57.207Z on GitLab

After a discussion with @basnijholt, we think that maybe printing a warning.warn(...) may be preferred over an error, and then just to disregard any point that is added outside of the domain.

[basnijholt]

originally posted by Jorn Hoofwijk (@Jorn) at 2018-09-20T10:51:01.390Z on GitLab

After another discussion, we came up with the following: Add the point to the self.data dictionary, but do not add it to any other internal state (such as the loss dict etc.)

apart from that, we could decide to give a warning, or not: but only once by the following filter:

warnings.simplefilter('once', MyDeprecationWarning)
for _ in range(100):
    warnings.warn('Warning!', MyDeprecationWarning)

[basnijholt]

originally posted by Jorn Hoofwijk (@Jorn) at 2018-09-20T14:17:02.960Z on GitLab

on closer inspection, this would work for the learner1D and the learnerND, the learner2D however, takes into consideration all points that are in self.data:

self.ask() is calculated based on self.losses()

self.losses() is calculated based on self.ip()

self.ip() is calculated based on self.data

therefore if we add the x that is outside the domain into self.data it will be taken into account whenever we request a new point -> so we can't guarantee that self.ask() returns a point inside the domain if we have added points outside the domain to self.data for the learner2D

[basnijholt]

originally posted by Jorn Hoofwijk (@Jorn) at 2018-09-20T14:19:54.976Z on GitLab

this doesn't mean the idea is bad, it just means that it would require some extra thought for the Learner2D (probably changing the code of the Learner2D)

[basnijholt]

originally posted by Jorn Hoofwijk (@Jorn) at 2018-09-20T14:27:09.221Z on GitLab

in fact: one of the things that should be tested, is the following: (also for the other learners of course)

import adaptive
l = adaptive.Learner1D(lambda x: x, (-1, 1))

l.tell(-1, -1)
l.tell(1, 1)
l.tell(5, 5)

xs, _ = l.ask(1)
x, = xs
assert -1 <= x <= 1