Fix hybrid bug, double performance

[dellaert]

I could not really separate the functional pruning PR from this one. Here's everything in this PR:

Fixing the bug

• BIG: fix bug in HybridConditional::errorTree and HGFG::errorTree
• made errorTree much faster by calling polymorphic errorTree and using AlgebraicDecisionTree operators

discretePosterior

• renamed both HBN::evaluate and HGFG::probPrime that take VectorValues to discretePosterior: these are very powerful.

Simplified API

• removed HBN::logProbability: errorTree is all we need and this one has confusing semantics
• removed HGC::logProbability that takes continuous values: very problematic name and only used in one place: again, errorTree is all we need.

Functional pruning

• pruning is now functional
• got rid of duplicate (unused) prunerFunc
• added a write-access conditional to allow iSAM versions to be still imperative

Update Oct 1:

• Reworked HGC::prune to use ADT::max
• Rewrote HybridBayesNet::prune completely

Looking at pruned trees for testHybridEstimation is very informative: it seems that the ordering of the keys is working against us in terms of pruning: we have large duplicate trees:

step 2:

step 5:

...

step 14:

[dellaert]

Note: “doubling performance” refers to testHybridEstimation. Could be twice as fast because now the discrete factors are properly taken into account when pruning.

[varunagrawal]

Some questions, especially about a copy.

[varunagrawal]

Why a copy if everything is functional?

[dellaert]

You're right, it was left-over ! I'll fix

[dellaert]

Actually, I messed up the discrete part. Did you ever profile the pruning as it is? It's an exponential operation - that alone might be responsible for slowness in some experiments.

I'll fix but we have to chat about this more I think.

[varunagrawal]

I haven't. I would like to see your perf-profiling setup and make it standard across projects. :) Seems to be very useful.

[dellaert]

I'm on a mac. I just build the binary and use Instruments CPU profiler.

[dellaert]

OK, I fixed it.

@varunagrawal :

• You probably had pruning tests in your project, but I added some here so I could refactor HGC::prune based on ADT::max.
• You might want to check out his branch and try whether your experiments still work before we merge?

[dellaert]

@varunagrawal please merge this after you established your experiments work?

[varunagrawal]

This looks great so merging!