Add a "balancing" learner

[jbweston]

(original issue on GitLab)

opened by Anton Akhmerov (@anton-akhmerov) at 2017-07-26T17:20:45.853Z

This should be an indispensable tool for parallel adaptive workflow: a container storing several similar learners and minimizing the loss that is a sum of the individual learner losses. So basically it should be a relay requesting points from the learner expecting highest immediate improvement.

[jbweston]

originally posted by Anton Akhmerov (@anton-akhmerov) at 2017-07-26T17:44:10.453Z on GitLab

Actually this one is an important test of our proposed structure.

In the current codebase, the function that is being learned is not visible to the learner which makes sense.

However in the current implementation of run this balancing learner would require the user to manually compose a function that has a arguments the number of the function that's being evaluated and the extra arguments, selected the correct function from a list, and plugs the extra arguments into it.

Should we move f to learner API? Should we implement balancing not as a learner? Should we let the user live with the boilerplate?

[jbweston]

originally posted by Joseph Weston (@jbweston) at 2017-08-16T13:19:02.305Z on GitLab

N.B. Round-robin is a good balancing strategy.