Control the BFGS optimizer.
Usage
controlBFGS(
startingValues = "est",
initialHessian = ifelse(all(startingValues == "est"), "lavaan", "compute"),
saveDetails = FALSE,
stepSize = 0.9,
sigma = 1e-05,
gamma = 0,
maxIterOut = 1000,
maxIterIn = 1000,
maxIterLine = 500,
breakOuter = 1e-08,
breakInner = 1e-10,
convergenceCriterion = 0,
verbose = 0,
nCores = 1
)Arguments
- startingValues
option to provide initial starting values. Only used for the first lambda. Three options are supported. Setting to "est" will use the estimates from the lavaan model object. Setting to "start" will use the starting values of the lavaan model. Finally, a labeled vector with parameter values can be passed to the function which will then be used as starting values.
- initialHessian
option to provide an initial Hessian to the optimizer. Must have row and column names corresponding to the parameter labels. use getLavaanParameters(lavaanModel) to see those labels. If set to "gradNorm", the maximum of the gradients at the starting values times the stepSize will be used. This is adapted from Optim.jl https://github.com/JuliaNLSolvers/Optim.jl/blob/f43e6084aacf2dabb2b142952acd3fbb0e268439/src/multivariate/solvers/first_order/bfgs.jl#L104 If set to a single value, a diagonal matrix with the single value along the diagonal will be used. The default is "lavaan" which extracts the Hessian from the lavaanModel. This Hessian will typically deviate from that of the internal SEM represenation of lessSEM (due to the transformation of the variances), but works quite well in practice.
- saveDetails
when set to TRUE, additional details about the individual models are save. Currently, this are the Hessian and the implied means and covariances. Note: This may take a lot of memory!
- stepSize
Initial stepSize of the outer iteration (theta_next = theta_previous + stepSize * Stepdirection)
- sigma
only relevant when lineSearch = 'GLMNET'. Controls the sigma parameter in Yuan, G.-X., Ho, C.-H., & Lin, C.-J. (2012). An improved GLMNET for l1-regularized logistic regression. The Journal of Machine Learning Research, 13, 1999–2030. https://doi.org/10.1145/2020408.2020421.
- gamma
Controls the gamma parameter in Yuan, G.-X., Ho, C.-H., & Lin, C.-J. (2012). An improved GLMNET for l1-regularized logistic regression. The Journal of Machine Learning Research, 13, 1999–2030. https://doi.org/10.1145/2020408.2020421. Defaults to 0.
- maxIterOut
Maximal number of outer iterations
- maxIterIn
Maximal number of inner iterations
- maxIterLine
Maximal number of iterations for the line search procedure
- breakOuter
Stopping criterion for outer iterations
- breakInner
Stopping criterion for inner iterations
- convergenceCriterion
which convergence criterion should be used for the outer iterations? possible are 0 = GLMNET, 1 = fitChange, 2 = gradients. Note that in case of gradients and GLMNET, we divide the gradients (and the Hessian) of the log-Likelihood by N as it would otherwise be considerably more difficult for larger sample sizes to reach the convergence criteria.
- verbose
0 prints no additional information, > 0 prints GLMNET iterations
- nCores
number of core to use. Multi-core support is provided by RcppParallel and only supported for SEM, not for general purpose optimization.