Here you can find out more about pdfbox and download it. Note: pdfbox is not yet available online. Contact me if you would like a distribution or more information…
FEATURES
automatic standard deviation estimation to improve error estimate (model and observation errors joined into one error term)
automatic lag-1 autocorrelation coefficient estimation (turned off / set to zero if not desired)
Observation error can be used to perform mcmc as normal (no sd)
Performs DE best fit optimization for comparison and/or mcmc starting point
A variable model time-step is allowed, with as many different time-steps as desired
MCMC assimilation can be completed with multiple observation sets, with sd, and ac estimation taking place for each observation set independently.
Allows for model constraints to easily be added.
Multiple confidence interval cut-off calculations are possible, specified by the user.
Simulations can be completed completely in R (slow) or in FORTRAN (very fast) to allow for flexibility, testing needs, etc.
Batch simulations are possible, easily sent to a queue
Allows for easy input of various data sources (file formats, etc.) into R, which can then be preprocessed and analyzed before sending to FORTRAN.
Formatting for FORTRAN input happens automatically – no user modifications should be needed.
Functions for plotting, visualization of output are included.
Code is generated for specific runs into a model directory. This directory can operate on it’s own, producing data, chains, etc. – useful for allowing others to reproduce code without needing entire toolbox…
LIMITS OF APPLICABILITY
For autocorrelation estimation to work properly, an obs. set must be sequential (no gaps, evenly-spaced) in the x-variable.
Standard deviation estimation does not take heteroscedasticity into account – for each obs. set, the sd parameter value will be set to the error at each x: error(:) = p.value
Linear interpolation of model to fit the data, handling of end-points is crude at best.
FUNCTIONAL NOTES
Make model output over observation interval finer or equal to the resolution of the obs. set.
Make model output equal to or longer (past and future) than the x-range of all obs. sets.
Alternatively, trim the beginning and end of the obs. set to x-range of model output, but this is less ideal.
FUTURE DEVELOPMENT IDEAS
Make modifications to the MCMC routine for improved performance (edge effect, optimized step-size determination, variable step-size, etc.)
Call FORTRAN model and constraints subroutines from R, to run R version faster (currently R compiles subroutines as FORTRAN 77 though, and FORTRAN90 is preferred – due to array allocation, etc. – workarounds may be possible).
Parallelize it!
Take heteroscedasticity of error into account (add/multiply factor onto obs. error, etc.).
Make autocorrelation estimation work with variable dt observations.
Comments are closed for this entry.