Fix cut sharing in a graph with zero-probability arcs

[odow]

Closes #796

@adow031 want to test this branch?

TODO

• add tests

[odow]

It'd be helpful if I could spell

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 93.58%. Comparing base (cc604a4) to head (68b40de).
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #797      +/-   ##
==========================================
+ Coverage   93.54%   93.58%   +0.03%     
==========================================
  Files          26       26              
  Lines        3519     3521       +2     
==========================================
+ Hits         3292     3295       +3     
+ Misses        227      226       -1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

---

🚨 Try these New Features:

• Flaky Tests Detection - Detect and resolve failed and flaky tests
• JS Bundle Analysis - Avoid shipping oversized bundles

[adow031]

This issue appears to be resolved.

But looking at the fix: is the result that the there's no longer cut sharing when there are some 0 probabilities? I would prefer that the code just inefficiently solved children with 0 probabilities than didn't share cuts.

(A user would always have the option to not include an arc with 0 probability if there's a preference for reducing the number of solves.)

[odow]

Currently we share cuts iff they have the identical set of children. We could improve that to nodes with have a subset of children. So yes, this might result in a less efficient solve; probably somewhere in between setting refine_at_similar_nodes = false, depending on the structure of the zeros.

[odow]

But yeah, I'll take another go at fixing this.

[odow]

Okay, I'm taking a look at this.

The issue is

SDDP.jl/src/algorithm.jl

Lines 755 to 757 in f921657

	if isapprox(child.probability, 0.0; atol = 1e-6)
	continue
	end

which we added in #224

The fix of just adding this is surprisingly subtle, because risk measures now need to deal with values that have zero probability mass

WorstCase is the obvious one, but it already checks this:

SDDP.jl/src/plugins/risk_measures.jl

Lines 49 to 57 in f921657

	for (index, (probability, observation)) in
	enumerate(zip(original_probability, objective_realizations))
	if probability > 0.0
	if (is_minimization && observation > worst_observation) ||
	(!is_minimization && observation < worst_observation)
	worst_index = index
	worst_observation = observation
	end
	end

[odow]

Belief states broke. I assume because something weird happens if you try to update the belief from somewhere that it was impossible to come from. There's a subtle distinction between ambiguity sets and nodes with similar children...

[odow]

Bump @adow031 interested in your thoughts