Possible recommenders

[nbgl]

• Greedy uncertainty
• Point that minimises sum of variances of test samples
• Point that minimises sum of squared errors of test samples (assuming y=μ)
• Most change of the function (assuming y=μ)
• Hyperopt

@chengsoonong, anything to add?

[chengsoonong]

Here's a potentially useful paper that summarises a few approaches, but only focuses on one for the rest.
https://arxiv.org/pdf/1206.2944.pdf
Also don't forget the four frameworks for active learning in the acton repo, issue 93-96
https://github.com/chengsoonong/acton/issues

[nbgl]

Cool, I’ll add Hyperopt to the list above

[nbgl]

My modification of UCB appears to work reasonably well, depending on the choice of the parameter k. Lower k values improve learning at the start, but work poorly towards the end of the learning curve. Higher k values learn poorly at the start, but very well towards the end. Note that uncertainty recommendation is the limiting case as _k_⟶∞.

I have been attempting to interpolate, with reasonable results. Still needs more tuning.

[nbgl]

^ This is slightly better interpolation, based on the tan function.

[chengsoonong]

One way to choose the schedule for changing k is to use sqrt(n).
See:
https://arxiv.org/pdf/0912.3995.pdf

[nbgl]

I need to write this up in thesis.