ENH: Updates to KPI options

[flongford]

Implements enhancements that take into account to updates to both KPISpecification and BaseOptimizerEngine classes introduced in force-h2020/force-bdss#353

The IOptimizer classes now don't have to deal with converting the raw KPI values into the objective score - all of this is dealt with by BaseOptimizerEngine._score

To Do:

• Include handling of upper_bound argument for NevergradMultiOptimizer class. In particular we need to use appropriate Nevergrad defaults for any KPI that does not have an upper bound set. I am not quite sure what these should be, but this previous issue may help: Bounded numerical value with shifted prior facebookresearch/nevergrad#361

Notes:

• It does not seem appropriate to use 'infinite' upper bounds for any KPI. Ideally the user should always specify upper bounds.

• Nevergrad's MultiobjectiveFunction currently deals with non-specified upper_bounds attributes by performing a few non-biased calculations of the objective function and identifying the largest values that are generated for each output variable (https://github.com/facebookresearch/nevergrad/blob/master/nevergrad/functions/multiobjective/core.py#L51).

• This is an 'all-or-nothing' approach, however - it will only perform this estimation if no upper_bounds array is passed into the constructor. Therefore it cannot handle certain KPIs with upper bounds set and others without (as currently allowed by the BDSS framework)

[codecov]

Codecov Report

Merging #23 into master will increase coverage by 0.59%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master      #23      +/-   ##
==========================================
+ Coverage   98.47%   99.06%   +0.59%     
==========================================
  Files           7        7              
  Lines         197      215      +18     
  Branches       32       34       +2     
==========================================
+ Hits          194      213      +19     
  Misses          2        2              
+ Partials        1        0       -1     

Impacted Files	Coverage Δ
force_nevergrad/engine/parameter_translation.py	96.36% <ø> (-0.07%)	⬇️
force_nevergrad/engine/nevergrad_optimizers.py	100.00% <100.00%> (+1.85%)	⬆️
force_nevergrad/mco/ng_mco.py	100.00% <100.00%> (ø)
force_nevergrad/mco/ng_mco_model.py	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update e4024d8...09bde2d. Read the comment docs.

[sparsonslab]

I've played around with it on the two-Gaussian example, using various lower/upper ranges (negative, positive, narrow, wide). The engine is still running and the Pareto-front is where is should be, certainly with a wide range. I'm not sure in theory what behaviour it should cause in other cases but it does seem to be having some effects.