Problem Interface

[aml5600]

• creates an AbstractUncertaintyProblem abstract type and ExpectationProblem struct
• adds solve methods for ExpectationProblem using Koopman and MonteCarlo
• moves organizational logic out of solve and into problem construction
• makes organizational logic more general...should be able to mix multivariate, univariate, and scalar distributions/values within the state and parameter arrays
• added uncertainty_utils.jl for helper functions

Relevant tests pass...The quadrature methods do not seem to work with ForwardDiff as the dual variables "enter" into the problem's lower/upper bounds. The construction of the problem should make it easy to setup for various problem types, this one should handle what I would like it to do at the moment!

[ChrisRackauckas]

DiffEqBase.

[aml5600]

can you elaborate?

[ChrisRackauckas]

DiffEqBase.solve

[aml5600]

gotcha, thanks!

[aml5600]

@ChrisRackauckas this causes HCubatureJL to fail for nout=1...if I switch the integrand to oop, it passes....

I dug into it, and it executes the exact same quadrature evaluation points and returning integrand values for both directly returning a scalar and in-place for a length=1 array. The quadrature alg seems to keep evaluating more points for the latter therefore returning a different value.

Any ideas?
FYI @agerlach

[agerlach]

HCubature will internally call different algorithms depending on if scalar bounds or array. This may explain it??? See hcubature vs hquadrature https://github.com/JuliaMath/HCubature.jl

[aml5600]

So I cannot get ForwardDiff to work with anything but HCubatueJL and that is with manually setting the upper and lower bounds. If the dual variable enters in through there then it breaks on HCubtaure.jl line 127, it seems this is a known problem (see here ).

I would imagine that the only reason the original tests pass is because the AD does not enter the bounds via the specific problem we are testing.

On the bright side, Zygote has no problems with the "full AD" problem!

[ChrisRackauckas]

So I cannot get ForwardDiff to work with anything but HCubatueJL and that is with manually setting the upper and lower bounds. If the dual variable enters in through there then it breaks on HCubtaure.jl line 127, it seems this is a known problem (see here ).

HCubature should get a PR to fix this. I think we can merge without that, but the tests should be set to properly mark which are broken and which are not (and in the current setup that will require some conditionals)

[agerlach]

@aml5600 can you update the readme to reflect the interface change? Also, SciMLTutorials SciML/SciMLTutorials.jl#343

How does or will this interface work w/ convenience methods like centralmoments()