Skip to contents

Train model

trainSWR.Rd

Trains an SWR model based on input and target time series data.

Usage

trainSWR(
  ts_input,
  ts_output,
  iter = 5,
  runs,
  log = FALSE,
  parallel,
  return = "best",
  param_selection = "best_bic",
  algorithm = "GENOUD"
)

Arguments

ts_input

a vector or ts object containing the input time series

ts_output

a vector or ts object (on the same time scale as ts_input) containing the target time series

iter

number of iterations (maximum number of windows)

runs

[Deprecated] number of independent model runs; no longer supported due to deterministic window initialization

log

whether a log-linear model should be used

parallel

[Deprecated] should the runs be computed in parallel? If FALSE, all runs are computed in serial. If TRUE, all runs are computed in parallel with a maximum number of cores. If a scalar is provided, the number of cores is set manually. No longer supported for single-run models

return

[Deprecated] either "best" (best model run is returned), or "all" (all model runs are returned)

param_selection

either "max" (maximum number of windows), or "best_rmse", "best_aic", or "best_bic" to optimize RMSE, AIC, or BIC, respectively

algorithm

either "GENOUD" (genetic optimization using derivatives), or "BOBYQA" (bound optimization by quadratic approximation)

Value

an object of type SWR model

Details

The training procedure implements an iterative algorithm described in (Schrunner et al. 2023) . A new window is added in each iteration, hence the number of windows equals to the iteration counter. Input and output time series are provided in ts_input and ts_output, respectively. Both are required to have equal lengths. The optimization is performed using the GENOUD algorithm (Mebane Jr. and Sekhon 2011) , implemented in rgenoud (Mebane Jr. and Sekhon 2023) , or alternatively using the BOBYQA algorithm (Powell 2009) , implemented in nloptr (Johnson 2021) . As training hyperparameter, iter indicates the number of iterations, which equals to the maximum number of windows selected by the model. Parameters return and param_selection indicate which criterion should be used to determine the number of windows; option are: return="all" (no hyperparameter selection) or return="best", which allows one of the following options:

  • param_selection="best_aic": select model with lowest AIC,

  • param_selection="best_bic": select model with lowest BIC,

  • param_selection="best_rmse": select model with lowest RMSE. Arguments runs, parallel are deprecated and should not be used.

References

Johnson SG (2021). The NLopt nonlinear-optimization package. v2.7.1, http://github.com/stevengj/nlopt.

Mebane Jr. W, Sekhon JS (2011). “Genetic Optimization Using Derivatives: The rgenoud package for R.” Journal of Statistical Software, 42(11): 1-26., 42(11).

Mebane Jr. W, Sekhon JS (2023). R Version of GENetic Optimization Using Derivatives. v5.9-0.10, https://github.com/JasjeetSekhon/rgenoud.

Powell MJ (2009). “The BOBYQA algorithm for bound constrained optimization without derivatives.” Cambridge NA Report NA2009/06, University of Cambridge, Cambridge, 26.

Schrunner S, Janssen J, Jenul A, Cao J, Ameli AA, Welch WJ (2023). “A Gaussian Sliding Windows Regression Model for Hydrological Inference.” arXiv.org (preprint). 2306.00453.

Examples

# train a model based on one year of observations
set.seed(42)
data(sampleWatershed)
mod <- trainSWR(sampleWatershed$rain[1:365],
                sampleWatershed$gauge[1:365],
                iter = 2)
#> Warning: The `runs` argument of `trainSWR()` is deprecated as of SlidingWindowReg 0.1.1.
#> ℹ Multiple model runs are not useful with deterministic window initializations.
#> Warning: The `parallel` argument of `trainSWR()` is deprecated as of SlidingWindowReg
#> 0.1.1.
#> ℹ Parallel model runs are not useful with deterministic window initializations.
#> Warning: The `return` argument of `trainSWR()` is deprecated as of SlidingWindowReg
#> 0.1.1.
#> ℹ Return argument is unused.
summary(mod)
#> SlidingWindowReg (SWR) model object with k = 2 windows
#> 
#> | window| delta| sigma| beta|
#> |------:|-----:|-----:|----:|
#> |      1|  0.17| 21.28| 0.36|
#> |      2|  0.86|  0.36| 0.67|
#>