Robustify BPS numerical machinery

[alexandrebouchard]

Latest failure example:

     --model.plate.maxSize 100 \
     --model.diagonal false \
     --engine PT \
     --engine.ladder.nChains 1 \
     --engine.nThreads 1 \
     --engine.random 1 \
     --engine.nPassesPerScan 1 \
     --samplers.useAnnotation false \
     --samplers.additional blang.mcmc.BouncyParticleSampler

More generally, should rewrite the numerical solver, make the numerical differentiator smarter, etc.

[alexandrebouchard]

Fails on first gradient eval. Cause: large values. Solutions: use relative error, also, monitor quality of approx, and when it starts to degrade back off.

[alexandrebouchard]

Comparing with BPSFactory on same data

40 dim blang generated data

wallClockTimeMillis	2325
trajectoryLength	1000.0
nJumps	124946
nQueuePolls	3210216

executionMilliseconds : 2415
outputFolder : /Users/bouchard/w/rejectfree/results/all/2018-01-05-06-36-12-dA9VK13b.exec

from Blang

nJumps	3928544
nQueuePolls	3928544
executionMilliseconds : 26188

[alexandrebouchard]

Blang in all one traj

nJumps  3852108
millis 24194

[alexandrebouchard]

Could it be because of fragmentation of the Poisson?

[alexandrebouchard]

With allInOnePoisson (all in one traj) + BRENT

nJumps 284175
millis 132094