User Tools

Site Tools


start:hype_file_reference:optpar.txt

This is an old revision of the document!


optpar.txt

The file holds additional model options to info.txt and is therefore located in the same folder as info.txt. The file is used to define what kind of optimisation to be done if calibration is set in info.txt. There are several different methods to chose from, each with their settings. Which model parameters to calibrate and within which boundaries are information also given in optpar.txt.

Maximum 100 model parameters may be optimised simultaneously. To optimise more parameters, the code needs to be changed (set maxoptpar to a higher value). All parameters are described in the section on par.txt, but not all of them can be calibrated. The objective function of the optimization is defined in info.txt as the combination of criteria chosen, see Performance criteria options.

There are eight methods of optimisation implemented in HYPE as detailed in the table below. Additionally, there are two other tasks for output generation, WA and WS, which produce detailed performance and simulation results for all runs performed during optimisation. Tasks WA and WS are compatible with selected optimisation methods only, a denoted in the table. The tast of organized scanning SC is a parameter investigation method.

Task Description
MCMonte Carlo (MC) simulation with parameter values randomly distributed over the intervals (basic MC-method)
BPprogressive Monte Carlo simulation with parameter space limited by best found so far (alternative MC-method)
SMprogressive Monte Carlo simulation with parameter space reduced in stages (alternative MC-method)
DEdifferential evolution Markov Chain method (alternative MC-method). Note
BNoptimisation with Brent method
Q1optimisation with QuasiNewton DFP gradient-based method
Q2optimisation with QuasiNewton BFGS gradient-based method
SDoptimisation with steepest descent method
WAwrite performance result for all simulations (MC, SM or DE)
WSwrite simulation results (basin-, time- or map-files) for all ensembles in Monte Carlo simulation (MC, BP, or DE) (maximum 9999999 ensembles total)
SCorganised scanning of two parameters

File content

The first row is for comments. It is ignored by the program. Second to 21st row are used to define tasks and calibration settings. Parameters to be calibrated are defined from row 22 and onward.

A row starts with a code indicating task or other settings. Argument of the code is listed from character 12 and forward on each row. The following options are available:

Code Argument Description
tasktwo letter worddefine what kind of optimisation to do (see methods above), and if all results are to be written for the MC-methods
cal_logY/Nflag for writing calibration.log file, default is Y
scan_numxintegernumber of steps taken for the first parameter (SC method)
scan_numyintegernumber of steps taken for the second parameter (SC method)
num_mcintegernumber of Monte Carlo simulations (per centre point and stage for progressive MC)
num_ensintegernumber of best Monte Carlo simulations to keep and print results from (and use as centre points for next stage of progressive MC) (maximum 999)
num_bpmcintegernumber of simulations per reduced parameter space which the best simulations shall be selected from (BP MC-method)
num_bpmaxintegernumber of reductions of the parameter space for MC simulation (BP MC-method)
num_stagesintegernumber of stages for progressive Monte Carlo (SM MC-method)
num_zoomrealreduction of parameter space (0-1) for each stage of progressive MC (BP MC-method)
DEMC_ngenintegernumber of generations for DE method
DEMC_npopintegernumber of populations for DE method
DEMC_gammascalerealscaling of the mutation strength for DE method. A new (next generation) parameter candidate is proposed as a mutation of the parent parameter value based on the difference between two random members of the parent population. DEMC_gammascale is a scaling factor for the resulting parameter jump width. Small values will cause smaller mutations, which potentially stabilises the search through large parameter spaces at the cost of convergence speed. A value of 1 will result in no scaling.
DEMC_crossoverrealcrossover probability for DE method. Probability that the proposed candidate is chosen instead of the parent parameter. Large DEMC_crossover values mean larger probability that the proposal is chosen. Set to 1, all proposals are accepted. This makes it harder to find an acceptable overall proposal because all parameters are changed in every generation. Set to 0.5, each parameter candidate has only a 50% chance to be accepted into the next proposal.
DEMC_sigmarealsample error standard deviation for DE method. Base for the standard deviation of the random perturbation, which adds random noise to the proposed parameter in addition to the gamma-mutation. This value is multiplied with 3rd-row value for each parameter (see description of parameter rows below).
DEMC_accprobintegerscaling factor for probabilistic acceptance for DE method (0 = off (default); >0 = on). If set to off, parameter proposals will only (and always) be accepted if the likelyhood score decreases (= better performance). Values close to 0 will increase the probability of acceptance within the probabilistic framework, a value of 1 means no scaling.
BR_diagStpY/Nflag for taking a diagonal step at the end of each iteration (BN method)
num_maxItrintegermax amount of iterations (interrupt non-MonteCarlo methods)
num_maxTimintegermax calibration time (hours) (interrupt non-MonteCarlo methods)
num_parItrintegernumber of iterations taken into account for parameter change tolerance (interrupt non-MonteCarlo methods)
num_criItrintegernumber of iterations taken into account for criterium change tolerance (interrupt non-MonteCarlo methods)
num_criTolrealtolerance for criteria relative change over last iterations (interrupt non-MonteCarlo methods)
lnS_maxItrintegermax amount of line search iterations (per line) (non-Monte Carlo methods)
lnS_tolrealgeneral relative tolerance for line search (non-Monte Carlo methods)
QN_nrmTolrealtolerance for gradient norm to be considered zero (QN methods)
QN_pctDervrealfactor to offset current parameter value for numerical derivative (QN methods)
QN_stencilintegernumerical derivative stencil type (2, 4, 6 and 8 allowed) (QN methods)
QN_lambMaxrealfactor to contain lambda (within given parameter intervals) prior to line search (QN methods)
QN_lambAccrealfactor increasing the step length (QN methods)

From row 22 and onward, parameters to calibrate are defined (no quotation marks). For class- or land use-dependent parameters, values for all soil-types/land-uses/subbasin/parameterregion/lakeregions or one value for general parameter have to be provided.

Each parameter is defined on three rows:

  • Row 1 specifies lower boundaries of the parameter range
  • Row 2 specifies upper boundaries of the parameter range (the model actually accepts lower and upper boundaries in any order)
  • Row 3 specifies either a minimum step width for parameter change or, in case of the DE method, a parameter specific additional factor to scale the random noise added to the proposed next-generation parameter, see description of DEMC settings code DEMC_sigma in table above.

NOTE: If lower and upper boundaries are identical, the parameter is omitted. This allows to calibrate a selection of class- or region-dependent parameters.

Example of parameter rows:

wcfc 	0.100 0.020 0.120 0.050 0.250 0.250 0.150 0.050 0.500 0.500 0.050
wcfc 	0.100 0.120 0.120 0.050 0.250 0.250 0.150 0.050 0.500 0.500 0.050
wcfc 	0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001 0.001
ttmp  2.0	0.0
ttmp  0.0	2.0
ttmp  0.01	0.01

The rows starting with wcfc are representing field capacity for 11 soil-types, where the second soil type's wcfc is calibrated. Parameter ttmp is threshold temperatures for 2 land-uses.

start/hype_file_reference/optpar.txt.1483457712.txt.gz ยท Last modified: 2023/11/16 14:28 (external edit)