minimize_objective: Calculate parameters than minimize objective function of...
In JGCRI/gcamland: GCAM Land Allocation Module

Description Usage Arguments Details Author(s)

Scans the runs in a given sample of model runs and selects the parameter set from those runs that minimizes a specified objective function across specified landtypes. The final minimized value is currently coded as the mean across specified landtypes of the specified objective function of interest.

minimize_objective(
  samples,
  modelgroup = "expectation.type",
  reportvars = NULL,
  objfun_to_min = "rms",
  landtypes = NULL
)

`samples`	Monte Carlo samples, given either as a grand table or a list of `ScenarioInfo` objects
`modelgroup`	Vector of names of columns that define the model groupings. The default is the single column `expectation.type`.
`reportvars`	Vector of names of variables for which to report expectations. The default is all parameter values.
`objfun_to_min`	a single objective function to minimize, from the supported c('bias', 'rms', 'centeredrms', 'nrms', 'centerednrms', 'kge')). Defaults to rms. Must be of length 1.
`landtypes`	Vector of land types to include in minimization. (default is to use `c( "Corn", "FiberCrop", "MiscCrop", "OilCrop", "OtherGrain", "PalmFruit", "Rice", "Root_Tuber","SugarCrop", "Wheat"))`.

For 'bias', the quantity being minimized is actually the average over specified land types of abs('bias'), as a bias of 0 is perfect performance, and a bias of 0.25 vs -0.25 are equally 'good'.

For 'kge', the quantity being minimized is actually the average over specified land types of 1 -'kge'. This is chosen because of the following:

Values of 'kge' can range anywhere from -infinity to +1, with a value of +1 being perfect model performance, therefore 'kge' is not directly useful as a quantity to be minimized.
Values of -'kge' range from -1 to +infinity, with -1 corresponding to perfect model performance, as well as being the minimum value achieved by this measure. However, because this quantity can be either negative or positive, it is possible that cancellation of errors may occur when calculating the mean across specified land types, leading to a false minimum.
Values of 1-'kge' range from 0 to +infinity, with 0 corresponding to perfect model performance, as well as being the minimum value achieved by this measure. Additionally, because all values of this quantity are positive, no cancelation of errors may occur when calculating the mean across specified land types.

ACS May 2020

JGCRI/gcamland documentation built on Oct. 6, 2020, 5:30 p.m.