Tools for non-intrusive and scalable parameter estimation and uncertainty quantification
PEST++ is a software suite aimed at supporting complex numerical models in the decision-support context. Much focus has been devoted to supporting environmental models (groundwater, surface water, etc) but these tools are readily applicable to any computer model.
builds and basic testing:
Links to latest binaries
Links to latest binaries
- windows (users with current visual studio installed). Direct zip download here
- windows compiled with intel C++ (the 'i' prefix). Direct zip download here
- mac OS. Direct zip download here
- linux static compiled with intel C++. Direct zip download here. (makefiles for GCC provides in src/ directory). These binaries are lagging behind until I can secure access to these compilers again...
The PEST++ software suite includes several stand-alone tools for model-independent (non-intrusive) computer model parameter estimation and uncertainty analysis. Codes include:
pestpp: deterministic GLM parameter estimation using "on-the-fly" subspace reparameterization, effectively reproducing the SVD-Assist methodology of PEST without any user intervention and FOSM-based parameter and (optional) forecast uncertainty estimation with support for generating posterior parameter realizations.
pestpp-gsa: Global senitivity analysis using either Morris or Sobol
pestpp-swp: a generic parallel run utility driven by a CSV file of parameter values
pestpp-opt: chance-constrainted linear programming
pestpp-ies: iterative ensemble smoother implementation of GLM (based on the work Chen and Oliver 2013) with support for generic localization (local analysis and/or covariance localization)
All members of the software suite can be compiled for PC, MAC, or Linux and have several run managers to support parallelization. precompiled binaries are available in the "bin" folder. Windows users with older OS versions should use the
bin/iwin binaries (starting "i", compiled with intel C++) to avoid the dreaded MSVC missing runtime DLL issue
update 12 November 2018 :
pestpp-ies now supports localization available as combined local analysis/covariance localization. This is controlled by a localization matrix which lists adjustable parameters and/or parameter groups as columns and non-zero weighted observations and/or observation groups as rows. The format of this matrix can be PEST ascii, PEST binary or CSV.
update 3 November 2018 : The FOSM calculations in
pestpp now also support generation and evaluation of a posterior parameter ensemble. If you add
++num_reals(50) to the end of the pest control file, once the GLM parameter estimation process of
pestpp is complete, during the FOSM parameter uncertainty calculations, 50 posterior parameter realizations are generated from the Schur-based (bayes linear) posterior parameter covariance matrix. These realizations are saved to a CSV file and are also evaluated (in parallel is
pestpp is running in parallel).
update 3 November 2018 : The repo structure has been significantly refactored in an effort to reduce the size the primary code and binaries. The main culprit was the
benchmarks directory, so the various benchmarks have now been split off into separate repos: https://github.com/jtwhite79/pestpp_benchmarks, https://github.com/jtwhite79/pestpp-ies_benchmarks, and https://github.com/jtwhite79/pestpp-opt_benchmarks. This makes the main repo (this repo) much smaller; the CI testing continues as before by using some tricks for automatically triggering downstream/dependent builds in travis - a commit against this repo will trigger a fake commit to each of those benchmark repos and subsuquently CI testing in travis and appveyor.
update 15 September 2018 : An official PEST++ V4 manual is now available in the Documentation directory. It is a docx file so any contributions (revisions/extensions/clarifications/etc) are greatly appreciated!
update 5 August 2018 : Welcome to the active fork of
pestpp-swp. I strive to actively support users, so please raise issues related to these codes as needed. I will also try to support
pestpp users are best I can although this code is a lower priority for me.
update 30 July 2018 : The PEST++ now supports parameter and observation names up to 200 characters in length. This allows for more descriptive naming and better support for problems with very large numbers of pars and obs. Note that if a parameter name exceeds 12 chars or an observation name exceeds 20 chars, the resulting jacobian binary file will include truncated names (at 12 and 20 chars, respectively) and names exceeding these lengths are not backward compatible with PEST.
update 4 July 2018 : PEST++ version 4.0.0 has been released to support the newly-developed
pestpp-ies. A manuscript documenting
pestpp-ies is available here: https://www.sciencedirect.com/science/article/pii/S1364815218302676. Stay tuned for an actual manual to accompany version 4!
update 2 May 2018 : some refactoring is underway.
sweep has been renamed
gsa has been renamed
pestpp-gsa. Also, the initial version of the new iterative ensemble smoother is avaiable as
pestpp-ies. The basic
++ options needed for fine-grained control of
pestpp-ies are listed below.
update 09/20/2017: the new optimization under uncertainty tool is ready! A supporting publication is in the works and should be available soon (a link will be posted once it is accepted). This new tool uses the same control file/template file/instruction file approach as other PEST(++) applications, so applying this tool to your problem should be seamless. Optional "++" args for tool are available further done this page.
update 01/25/2017: intel C++ builds are avaiable for mac and for windows. For mac users, these are statically-linked so they do not require compilers to be installed. For windows users, the intel build circumvents the "missing VCOMP140.DLL" error. Note the intel windows builds are currently in the
update 11/25/2016: PEST++ version 3.6 is now available. Some of the many enhancements available in 3.6 include:
a new approach to implementing regularization. Rather than using the standard pest control file parameters such as
fracphim, etc, we now offer a single pest++ argument,
++reg_frac(), that allows users to specify what fraction of the composite objective function should be regularization penalty. For example,
++reg_frac(0.5)would result in equal parts data misfit and regularization penalty, which results in the maximum a posteriori (MAP) parameter estimate. Using
++reg_frac()will result in substantial speed ups during the lambda calculation process
a new program for sequential linear programming under uncertainty.
pestpp-optis a new executable in the PEST++ suite that uses the standard PEST model independent interface to solve a (sequential) linear programming (LP) problem.
pestpp-optrelies on the COIN-OR Linear Programming (CLP) solver https://projects.coin-or.org/Clp. Also, users have the option to use FOSM-based uncertainty estimation in the evaluation of model-based constraints (such as water levels, stream flows, stream flow depletion, etc) so that a risk-based optimal solution can be found. See below for the required and optional
++arguments needed to apply
pestpp-opt. Two example problems using the
pestpp-opttool have been added to the
benchmarksdir. A publication about this tool is in the works.
global optimization with differential evolution. We now have a fully-parallel global solver that implements the differential evolution algorithm (DE) integrated into the pest++ executable. See below for required and optional
++arguments needed to use the DE solver.
a new randomization-based SVD solver using the implementation of https://github.com/ntessore/redsvd-h. This solver is activated using
++svd_pack(redsvd). Testing shows it to be very efficient for problems for a wide range of problem sizes, especially with judicious use of
upgrade parameter covariance scaling. Through the
++parcov_scale_fac(), pest++ can now scale the normal matrix (J^tQJ) by a user specified parameter covariance matrix. If no
++parcov_filename()is provided, pest++ will construct a diagonal parameter covariance matrix from the parameter bounds. This is a relatively new option and needs more testing, but limited testing to date shows that upgrade vectors resulting from a covariance-scaled normal matrix are more in harmony with expert knowledge.
Update 05/26/2016: PEST++ V3 has been officially released. It supports a number of really cool features, including global sensitivity analyses, and automatic Bayes linear (first-order, second-moment) parameter and forecast uncertainty estimates. We also have a utility for fully-parallel parametric sweeps from csv-based parameter files, which is useful for Monte Carlo, design of experiments, surrogate construction, etc. All of these tools are based on the model-independent communication framework of PEST, so if you have a problem already setup, these tools are ready for you!
Update 10/1/2014: recent stable versions of PEST++ implement dynamic regularization, full restart capabilities, additional options for formulating the normal equations, and an iterative SVD algorithm for very-large problems. Additionally the YAMR run manager has been improved to use threaded workers so that the master can more easily load balance.
White, J. T., 2018, A model-independent iterative ensemble smoother for efficient history-matching and uncertainty quantification in very high dimensions. Environmental Modelling & Software. 109. 10.1016/j.envsoft.2018.06.009. http://dx.doi.org/10.1016/j.envsoft.2018.06.009.
White, J. T., Fienen, M. N., Barlow, P. M., and Welter, D.E., 2017, A tool for efficient, model-independent management optimization under uncertainty. Environmental Modeling and Software. http://dx.doi.org/10.1016/j.envsoft.2017.11.019.
Welter, D.E., White, J.T., Hunt, R.J., and Doherty, J.E., 2015, Approaches in highly parameterized inversion— PEST++ Version 3, a Parameter ESTimation and uncertainty analysis software suite optimized for large environmental models: U.S. Geological Survey Techniques and Methods, book 7, chap. C12, 54 p., http://dx.doi.org/10.3133/tm7C12.
Welter, D.E., Doherty, J.E., Hunt, R.J., Muffels, C.T., Tonkin, M.J., and Schreüder, W.A., 2012, Approaches in highly parameterized inversion—PEST++, a Parameter ESTimation code optimized for large environmental models: U.S. Geological Survey Techniques and Methods, book 7, section C5, 47 p., available only at http://pubs.usgs.gov/tm/tm7c5.
The master branch includes a Visual Studio 2015 project, as well as makefiles for linux and mac. The suite has been succcessfully compiled with gcc (g++ and gfortran) 4,5,6 and 7 on ubuntu, fedora, and a slurm/MPI cluster - to use gcc, you need to have both
blas libraries available in the path.
benchmarks folder contains a simple worked example that is used for basic CI testing. Many full-worked test problems of varying problem sizes are now located in separate repos:
Much work has been done to avoid additional external dependencies in PEST++. As currently designed, the project is fully self-contained and statically linked.
blas are also required - these are included with the intel fortran compiler; those using gcc will need to have these libraries available.
please see the PEST++ version 4 manual in the
docs directory for a more complete description of these options
parallel run manager arguments
These are the optional
++ args that can be used to control the parallel run manager
++overdue_resched_fac(1.15):YAMR only, if a run is more than <
overdue_resched_fac> X average run time, reschedule it on available resources (depending on value of <
++overdue_giveup_fac(100.0):YAMR only, if a run is more than <
overdue_giveup_fac> X average run time, mark it as failed. Default of
++overdue_giveup_minutes(1.0e+30):YAMR only, if a run has been going for more than <
overdue_giveup_minutes>, mark it as failed.
++max_run_fail(4):maximum number of runs that can fail before the run manager emits an error and also the maximum number of concurrent runs allowed (for runs that are "overdue" according to <
overdue_resched_fac>). Default for
++condor_submit_file(pest.sub): a HTCondor submit file. Setting this arg results in use of a specialized version of the YAMR run manager where the
condor_submit()command is issued before the run manager starts, and, once a set of runs are complete, the workers are released and the
condor_rm()command is issued. This specialized run manager is useful for those sharing an HTCondor pool so that during the upgrade calculation process, all workers are released and during upgrade testing, only the required number workers are queued. As with all things PEST and PEST++, it is up to the user to make sure the relative paths between the location of the submit file, the control file and the instance of PEST++ are in sync.
Here is a (more or less) complete list of
++ arguments that can be added to the control file for
++max_n_super(20): maximum number of super parameters to use
++super_eigthres(1.0e-8)ratio of max to min singular values used to truncate the singular components when forming the super parameter problem
++n_iter_base(1):number of base (full) parameter iterations to complete as part of the on-the-fly combined base-parameter/super-parameter iteration process. A value of -1 results in calculation of the base jacobian and formation of the super parameter problem without any base parameter upgrades, replicating the behavior of the "svd-assist" methodology of PEST
++n_iter_super(4): number of super (reduced dimension) parameter iterations to complete as part of the on-the-fly combined base-parameter/super-parameter iteration process
++svd_pack(redsvd): which SVD solver to use. valid arguments are
propack(iterative Lanczos solution) and
++lambdas(0.1,1,10,100,1000): the values of lambda to test in the upgrade part of the solution process. Note that this base list is augmented with values bracketing the previous iterations best lambda. However, if a single value is specified, only one lambda will be used.
++lambda_scale_fac(0.9,0.8,0.7,0.5): the values to scale each lambda upgrade vector. This results in a line search along each upgrade vector direction, so that the number of upgrade vectors = len(lambdas) * len(lambda_scale_fac). To disable, set = 1.0.
++reg_frac(0.1): the portion of the composite phi that will be regularization. If this argument is specified, the
* regularizationsection of the control file is ignored. For limited testing, values ranging from 0.05 to 0.25 seem to work well.
++base_jacobian(filename): an existing binary jacobian file to use for the first iteration
++parcov(filename): an ASCII PEST-style matrix file or uncertainty file to use as the prior parameter covariance matrix in FOSM uncertainty calculations and/or normal matrix scaling. If not specified and a prior is needed, one is constructed on-the-fly from parameter bounds
++uncertainty(true):flag to activate or deactivate FOSM-based parameter and (optionally) forecast uncertainty estimation
++forecasts(fore1,fore2...):comma separated list of observations to treat as forecasts in the FOSM-based uncertainty estimation
++iteration_summary(true):flag to activate or deactivate writing iteration-based CSV files summarizing parameters (<base_case>.ipar), objective function (<base_case>.iobj) and sensitivities (<base_case>.isen), as well as upgrade summary (<base_case>.upg.csv) and a jacobian parameter-to-run_id mapping (<base_case>.rid).
++jac_scale(true): use PEST-style jacobian scaling. Important, but can be costly because it densifies the normal matrix, making SVD take longer.
++upgrade_augment(true): augment the values of lambda to test by including the best lambda from the previous iteration, as well as best lambda * 2.0 and best lambda / 2.0. If
true, then additional lambdas will be included by attempting to extend each upgrade vector along the region of parameter space defined by parameter bounds. If
false, then only the vectors listed in the
++lambda()arg will be tested and no extended upgrade will be included.
++upgrade_bounds(true): use additional tricks and upgrades to deal with upgrades that are going out bounds. If
true, can result in substantial phi improvement, but in some cases can produce NaNs in the upgrade vectors.
++hotstart_resfile(mycase.res): use an exising residual file to restart with an existing jacobian to forego the initial, base run and jump straight to upgrade calculations (++base_jacobian arg required).
++mat_inv(jtqj): the form of the normal matrix to use in the solution process. Valid values are "jtqj" and "q1/2j".
++der_forgive(true): a flag to tolerate run failures during the derivative calculation process
++parcov_scale_fac(0.01): scaling factor to scale the prior parameter covariance matrix by when scaling the normal matrix by the inverse of the prior parameter covariance matrix. If not specified, no scaling is undertaken; if specified,
++mat_invmust be "jtqj".
sweep is a utility to run a parametric sweep for a series of parameter values. Useful for things like monte carlo, design of experiment, etc. Designed to be used with
pyemu and the python pandas library.
++sweep_parameter_csv_file(filename): the CSV file that lists the runs to be evaluated. "sweep_in.csv" is the default
++sweep_output_csv_file(filename): the output CSV file from the parametric sweep. If not passed, output is written to "sweep_out.csv"
++sweep_chunk(500): number of runs to batch queue for the run manager. Each chunk is read, run and written as a single batch
++sweep_forgive(false): a flag to forgive missing parameters in the input csv file. If
true, then missing parameters are filled with the initial parameter value in the control file.
pestpp-opt is an implementation of sequential linear programming under uncertainty for the PEST-style model-independent interface
++opt_dec_var_groups(<group names>): comma-separated string identifying which parameter groups are to be treated as decision variables. If not passed, all adjustable parameters are treated as decision variables
++opt_external_dec_var_groups(<group_names>): comma-separated string identifying which parameter groups are to be treated as "external" decision variables, that is decision variables that do not influence model-based constraints and therefore do not require a perturbation run of the model to fill columns in the response matrix.
++opt_constraint_groups(<group names>): comma-separated string identifying which observation and prior information groups are to be treated as constraints. Groups for "less than" constraints must start with either "l_" or "less"; groups for "greater than" constraints must start with "g_" or "greater". If not passed, all observation and prior information groups that meet the naming rules will be treated as constraints
++opt_obj_func(<obj func name >): string identifying the prior information equation or two-column ASCII file that contains the objective function coefficients. If not passed, then each decision variable is given a coefficient of 1.0 in the objective function.
++opt_direction(<direction>): either "min" or "max", whether to minimize or maximize the objective function.
++opt_risk(<risk>): a float ranging from 0.0 to 1.0 that is the value to use in the FOSM uncertainty estimation for model-based constraints. a value of 0.5 is a "risk neutral" position and no FOSM measures are calculated. A value of 0.95 will seek a 95% risk averse solution, while a value of 0.05 will seek a 5% risk tolerant solution. See Wagner and Gorelick, 1987, Optimal groundwater quality management under parameter uncertainty for more background on chance-constrained linear programming
++opt_skip_final(<skip_final>): a flag to skip the final model run using optimal decision variable values. If
++hotstart_resfile()are set and (lot of "ands")
noptmax=0, then this causes no model runs to happen, pestpp-opt simply solves the chance constrainted LP problem, report optimal decision variables and phi, then exits. Default is
++opt_std_weights(<std_weights>):a flag to treat model-based constraints (listed in the
* observation datasection) as standard deviations for chance constraints. This can result in substantial time savings since the FOSM calculation process can be skipped. This can also be used to specific empirical constraint uncertainty (e.g. from ensemble methods). Default is
pestpp-ies is an implementation of the iterative ensemble smoother GLM algorithm of Chen and Oliver 2012. So far, this tool has performed very well across a range of problems. It functions without any additional
++ arguments. However, several
++ arguments can be used to fine-tune the function of
++ies_parameter_ensemble(<par_en>): file containing parameter ensemble. File extension is used to determine file type:
.jcbare supported. If not passed, a parameter ensemble is generated from prior parameter distribution
++ies_observation_ensemble(<obs_en>): file containing observation noise ensemble (obs vals + noise realizatons). File extension is used to determine file type:
.jcbare supported. If not passed, an observation ensemble is generated using observation weights
ies_restart_obs_en(<restart_obs_en>): file containting a restart observation ensemble (simulated outputs from a previous pestpp-ies run or from pestpp-swp). File extension is used to determine file type:
ies_num_reals(<num_reals>): number of realizations to use. If
restart_obs_enare passed and are larger than
num_reals, then only the first
num_realsrealizations are used. Default is 50.
ies_bad_phi(<bad_phi>): realizations yielding a phi greater than
bad_phiare dropped from
obs_en. Default is 1.0e+30
ies_lambda_mults(<lambda_mults>): lambda multiplers to test during upgrade calculations. Default is [0.1,0.5,1.0,2.0,5.0].
ies_initial_lambda(<init_lambda>): initial lambda value to use with
lambda_multsto get lambda values for testing during the first iteration. If not passed,
init_lambdais derived from the initial mean phi.
ies_subset_size(<subset_size>): the number of realizations to test for each upgrade parameter ensemble. The total number of upgrade testing runs is
subset_size. Default is 5, meaning only the first 5 realizations are evaluated during testing; if a successful (subset) upgrade ensemble is found, the remaining
subset_sizerealizations are evaluated.
ies_reg_factor(<reg_factor>): fraction of zeroth-order Tikhonov regularization penalty to add to the measurement phi to form composite phi. Default is 0.0
ies_use_approx(<use_approx>): flag to use the approximate upgrade solution process. Default is
ies_use_prior_scaling(<use_prior_scaling>): flag to scale various quantities by the prior parameter covariance matrix during upgrade calculations. Default is
ies_use_empirical_prior(<use_empirical_prior>): flag to calculate prior parameter covariance matrix from the parameter ensemble. Requires passing of
par_en. Default is
ies_verbose_level(<verbose_level>): integer to control how much pestpp-ies writes. Can be 0, 1, 2, or 3. Default is 1.
ies_add_bases(<add_bases>): flag to add initial parameter values to
par_enand add actual observation values (no noise) to
obs_en. This results in an approximation to the minimum error variance parameter set being carried through the pestpp-ies analysis. If
true, results in
num_reals+ 1 realizations. Default is
ies_enforce_bounds(enforce_bounds>): flag to enforce parameter bounds during upgrade calculations. Default is
ies_save_binary(<save_binary>): flag to save iteration parameter and observation ensembles to pest-compatible (jacobian format) binary files. Default is
ies_accept_phi_fac(<accept_phi_fac>): tolerance for accepting the results for a (subset) ensemble evaluation. If the resulting mean phi *
accept_phi_facis greater than the best mean phi from the last iteration, then the upgrade is rejected. Default is 1.05 (5% tolerance).
ies_lambda_inc_fac(<lambda_inc_fac>): factor increase current lambda by if current upgrade testing was not successful. Default is 10.0
ies_lambda_dec_fac(<lambda_dec_fac>): factor to decrease current lambd by if current upgrade testing was successful. Default is 0.75.
ies_save_lambda_en(<save_lambda_en>): flag to save lambda testing parameter ensembles. Can be use for finding parameters vectors that are causing run failures. Default is
parcov_filename(<parcov_filename>): (repeated argument for pestpp and pestpp-opt). Name of existing prior parameter covariance matrix. Can be ASCII (
.cov), binary (
.jcb) or an uncertainty file (
.unc). If not passed, the prior parameter covariance matrix is constructed from parameter bounds (or from the parameter ensemble if
par_sigma_range(<par_sigma_range>): the number of standard deviations implied by parameter bounds. Used to construct prior parameter covariance matrix from parameter bounds. Default is 4.0.
lambda_scale_fac(<lambda_scale_fac>): line search lambda scaling factors. Each lambda parameter upgrade ensemble is scaled by each
lambda_scale_fac. Default is [0.5,0.75,0.9,1.0,1.1].
ies_subset_how(<subset_how>): choice for how the subset is selected. Choices are "first" (use the first
subset_sizerealizations),"last" (use the last
subset_sizerealizations),"random" (randomly select
subset_sizerealizations each iteration),"phi_based" (chose
subset_sizerealizations spread across the composite phi from the last iteration). Default is "phi_based". If present, the
baseparameter realization will always be included in the subset.
ies_localizer(<localizer>): an optional localizer to use to localize spurious cross-correlations between observations and parameters. The file format is determined by the extension: "mat","cov","csv","jco"/"jcb". The row names of the matrix are observation names and/or observation group names and columns are parameter names and/or parameter group names (each obs/par can only be include once even through groups). Values should range between 0.0 and 1.0 (although no check is made). Adjustable parameters not listed in the columns are implicitly treated as
fixedand non-zero weight observations not listed in the rows are implicitly treated as zero-weighted.
ies_num_threads: the number of threads to use when calculating localized upgrades for large numbers of parameters. Default is
This software has been approved for release by the U.S. Geological Survey (USGS). Although the software has been subjected to rigorous review, the USGS reserves the right to update the software as needed pursuant to further analysis and review. No warranty, expressed or implied, is made by the USGS or the U.S. Government as to the functionality of the software and related material nor shall the fact of release constitute any such warranty. Furthermore, the software is released on condition that neither the USGS nor the U.S. Government shall be held liable for any damages resulting from its authorized or unauthorized use