Update parameter in Monte Carlo run

[grantmcdermott]

The package documentation provides a nice example of how to update a parameters using the !update_param API method. However, near as I can tell, this method only works for the deterministic version of the model(s). I'm looking for a way to update parameters in the Monte Carlo version of the model.

Using MimiPAGE2009.jl as an example, say I want to edit the TCR parameter (i.e. tcr_transientresponse here). I thought that adding the below line to the do_monte_carlo_runs() function might do the trick, but I just get a bunch of error messages:

update_param!(mcs, :tcr_transientresponse, Normal(0.0, 0.1)) ## Try crazy low TCR

Any suggestions for a better method are much appreciated.

[corakingdon]

Hi Grant, you are correct, the update_param! function is meant to be used on a deterministic model, not within the Monte Carlo simulation. If I am interpreting your intention correctly, you would like to sample very low values of TCR throughout the simulation (each trial with a different value from your normal distribution).

In order to do this, you should modify the definition for the tcr_transientresponse parameter's distribution that is definied within the @def_sim macro where all of PAGE's random variables are defined. On line 46 of "src/mcs.jl" in the MimiPAGE2009 package, the random variable is defined as the following:

tcr_transientresponse = TriangularDist(1, 2.8, 1.3)

Simply change this to your desired distribution

tcr_transientresponse = Normal(0.0, 0.1)

And then when you run the simulation (using the do_monte_carlo_runs function if desired) it will sample from this new distribution instead.

Let me know if you have any other questions about how to use this, I'm happy to help!

[grantmcdermott]

Super, thanks Cora. That approach was going to be my fallback, so it's good to have it sanity checked by someone in the know.

[grantmcdermott]

PS - It might be helpful to add a note to the docs about the limitations of !update_param to the deterministic case. Cheers!

[davidanthoff]

I think it would be good to come up with a more interactive API for this, though. I think in general if we have a situation where someone wants to run an alternative specification, we should make sure that we have a way to do this that does not involve changing a file from the original model package.

[grantmcdermott]

Yes, as a user that would be my strong preference too.

Similarly, I'm not entirely sure how to generate a SCC calculation using the stochastic framework at present. (See also #411). Say I follow this example and want to generate an emissions pulse of 1 billion metric tons of carbon in 2020. Do you have any suggestions for the best way to do this currently? (I'm specifically using MimiPAGE09 if that makes a difference.)

[corakingdon]

@davidanthoff Yes I agree, I've been saying I think we should add methods to modify a Simulation (such as add or change random variable distributions, and add or change which variables get saved)

@grantmcdermott I can get back to you in a bit about example code for SCC calculations in PAGE. The code doesn't currently exist in the MimiPAGE2009 package, but we are hoping to update with it soon

[lrennels]

Per our conversation today, we would like to add options to:

• add a random variable
• remove a random variable
• update a random variable
• update or change the save list

[rjplevin]

I've implemented much of this already, but some of it gets complicated. Modifying the save list is straightforward. The RVs are more complicated. Consider the example below.

Random variables can be named or anonymous. In the anonymous case, an Distribution is associated with a parameter by assignment, multiplication, or addition. In these cases, an internal RV is created from the variable name, "!" and an incrementing integer to ensure uniqueness, e.g., :depk!8. This is fully equivalent to the user defining this as rv(depk!8) = ... and then referencing depk!8 where the distribution is now given. It just saves some needless redundancy.

As originally conceived this approach with anonymous RVs was all "internal", but unless we want to redesign this, we will have to expose these generated names to the user to allow updates. Not a big deal, I suppose.

It's easy to support swapping out the RV for another, provided the name doesn't change. Deleting an RV should delete the TransformSpec to avoid referencing the defunct distribution. Adding an RV will require providing the TransformSpec as well.

julia> sim = @defsim begin
           # Define random variables. The rv() is required to disambiguate an
           # RV definition name = Dist(args...) from application of a distribution
           # to an external parameter. This makes the (less common) naming of an
           # RV slightly more burdensome, but it's only required when defining
           # correlations or sharing an RV across parameters.
           rv(name1) = Normal(1, 0.2)
           rv(name2) = Uniform(0.75, 1.25)
           rv(name3) = LogNormal(20, 4)
           # assign RVs to model Parameters
           share = Uniform(0.2, 0.8)
           sigma[:, Region1] *= name2
           sigma[2020:5:2050, (Region2, Region3)] *= Uniform(0.8, 1.2)
           depk = [Region1 => Uniform(0.08, 0.14),
                   Region2 => Uniform(0.10, 1.50),
                   Region3 => Uniform(0.10, 0.20)]
           sampling(LHSData, corrlist=[(:name1, :name2, 0.7), (:name1, :name3, 0.5)])
           
           # indicate which parameters to save for each model run. Specify
           # a parameter name or [later] some slice of its data, similar to the
           # assignment of RVs, above.
           save(grosseconomy.K, grosseconomy.YGROSS, emissions.E, emissions.E_Global)
       end
Simulation{LHSData}
  trials: 0
  current_trial: 0
  rvdict:
    name1: RandomVariable{Normal{Float64}}
    name2: RandomVariable{Uniform{Float64}}
    name3: RandomVariable{LogNormal{Float64}}
    share!6: RandomVariable{Uniform{Float64}}
    sigma!7: RandomVariable{Uniform{Float64}}
    depk!8: RandomVariable{Uniform{Float64}}
    depk!9: RandomVariable{Uniform{Float64}}
    depk!10: RandomVariable{Uniform{Float64}}
  translist:
    TransformSpec(:share, :(=), :share!6, Any[])
    TransformSpec(:sigma, :*=, :name2, Any[Colon(), :Region1])
    TransformSpec(:sigma, :*=, :sigma!7, Any[2020:5:2050, Symbol[:Region2, :Region3]])
    TransformSpec(:depk, :(=), :depk!8, Any[:Region1])
    TransformSpec(:depk, :(=), :depk!9, Any[:Region2])
    TransformSpec(:depk, :(=), :depk!10, Any[:Region3])
  savelist:
    (:grosseconomy, :K)
    (:grosseconomy, :YGROSS)
    (:emissions, :E)
    (:emissions, :E_Global)
  nt_type: NamedTuple{(:name1, :name2, :name3, :share!6, :sigma!7, :depk!8, :depk!9, :depk!10),NTuple{8,Float64}}
  0 models
  1 results dicts
LHSData(Tuple{Symbol,Symbol,Float64}[(:name1, :name2, 0.7), (:name1, :name3, 0.5)])