Extended IRPropMin with boundary support and methods to set min and max step size

[MichaelClerx]

⬆️ that

The boundary support is non-standard and can cause you to get stuck if you have a problematic problem (i.e. gradient next to the boundaries pointing outside). But very useful if you need to avoid evaluating outside some range that causes errors...

[codecov]

Codecov Report

Merging #1465 (d4ba352) into master (d72ca64) will not change coverage.
The diff coverage is 100.00%.

@@            Coverage Diff            @@
##            master     #1465   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files           97        97           
  Lines         9482      9532   +50     
=========================================
+ Hits          9482      9532   +50     

Impacted Files	Coverage Δ
pints/_optimisers/_irpropmin.py	100.00% <100.00%> (ø)

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

[MichaelClerx]

@DavAug maybe don't review just yet: I think i'll need to change the strategy for step-size shrinking with general boundaries

[MichaelClerx]

OK ready now @DavAug

• Added max step size
• Added get/set for max & min step size, although default min step size still set with sigma0
• Added boundary checks: for rectangular boundaries this is straightforward. For non-rectangular it's a bit weird: I first reduce all step sizes until the boundaries are OK. Then I try setting them all back to their original values again (one by one). The idea here is that if I need to reduce step sizes in dimensions 3, 7, and 8, then that will take me forever to find out if I start by trying to reduce just 1 step size, then all combinations of 2 step sizes, then all of 3. So this looks stupid but should work

[MichaelClerx]

(The strategy is biased against parameters/dimensions with a lower indice: these will get reduced first (even if reducing another parameter would also have worked). I thought a bit about randomisation but that makes this a stochastic method and I don't want to go there. The bias shouldn't matter, I hope: as long as we get to the right solution in the end)

[DavAug]

Thanks @MichaelClerx this looks very good! I have not reviewed everything yet, but I thought I leave some comments already. I'll have another look tomorrow 😊

[MichaelClerx]

Thanks!

[DavAug]

Looks good @MichaelClerx ! I also think that the step size adjustment for non-rectangular boundaries feels a little clumsy and, as you said, there is definitely bias to move towards smaller index dimensions. But at the same time, I think this will only start to matter once we introduce more complicated boundaries than currently available in pints.

[MichaelClerx]

Thanks @DavAug ! I've added the checks and merged it in