ProfileObjective {CollocInfer}R Documentation

Profile Estimation with Collocation Inference

Description

Profile estimation and objective functions for collocation estimation of parameters in continuous-time stochastic processes.

Usage

Profile.GausNewt(pars,times,data,coefs,lik,proc,in.meth="nlminb",
      control.in=NULL,active=1:length(pars),
      control=list(reltol=1e-6,maxit=50,maxtry=15,trace=1))

ProfileSSE(pars,allpars,times,data,coefs,lik,proc,in.meth='nlminb',
      control.in=NULL,active=1:length(pars),dcdp=NULL,oldpars=NULL,
      use.nls=TRUE,sgn=1)
      
ProfileErr(pars,allpars,times,data,coefs,lik,proc,in.meth = "house",
      control.in=NULL,sgn=1,active=1:length(allpars))

ProfileDP(pars,allpars,times,data,coefs,lik,proc,in.meth = "house",
      control.in=NULL,sgn=1,sumlik=1,active=1:length(allpars))

ProfileList(pars,allpars,times,data,coefs,lik,proc,in.meth = "house",
      control.in=NULL,sgn=1,active=1:length(allpars))

Arguments

pars

Initial values of parameters to be estimated processes.

allpars

Full parameter vector including pars. Assignment allpars[active] = pars is always made.

times

Vector observation times for the data.

data

Matrix of observed data values.

coefs

Vector giving the current estimate of the coefficients in the spline.

lik

lik object defining the observation process.

proc

proc object defining the state process.

in.meth

Inner optimization function currently one of 'nlminb', 'MaxNR', 'optim' or 'house'. The last calls SplineEst.NewtRaph. This is fast but has poor convergence.

control.in

Control object for inner optimization function.

sgn

Is the minimizing (1) or maximizing (0)?

active

Incides indicating which parameters of pars should be estimated; defaults to all of them.

oldpars

Starting parameter values for the Q-function in the EM algorithm.

dcdp

Estimate for the gradient of the optimized coefficients with respect to parameters; used internally.

use.nls

In ProfileSSE, is 'nls' being used in the outer-optimization?

sumlik

In ProfileDP and ProfileDP.AllPar; should the gradient be given for each observation, or summed over them?

control

A list giving control parameters for Newton-Raphson optimization. It should contain

  • reltol Relative tollerance criterion for the gradient and improvement before termination.

  • maxit Maximum number of iterations.

  • maxtry Maximum number of halving-steps to try before declaring no improvement is possible.

  • trace How much iteration history to output; 0 surpresses all output, a positive value outputs parameters and improvement at each iteration.

Details

Profile.GausNewt provides a simple implementation of Gauss-Newton optimization and may not result in optimized values that are as good as other optimizers in R.

When Profile.GausNewt is not being used, ProfileSEE and ProfileERR create the following temporary files:

These need to be removed manually when the optimization is finished. optcoefs.tmp will contain the optimal value of coefs for plotting the estimated trajectories.

Value

Profile.GausNewt

Output of a simple Gaus-Newton iteration to optimizing the objective function when the observation likelihood takes the form of a sum of squared errors. Returns a list with the following elements:

  • pars The optimized value of the parameters.

  • in.res The result of the inner optimization.

  • value The value of the optimized sum of squared errors.

ProfileSSE

Output for the outer optimization when the observation likelihood is given by squared error. This is a list with the following elements

  • value The value of the outer optimization criterion.

  • gradient The derivative of f with respect to pars.

  • coefs The optimized value of coefs for the current value of pars.

  • dcdp The derivative of the optimized value of coefs at the current value of pars.

ProfileErr

The outer optimization criterion in the general case: the value of the observation likelihood at the profiled estimates of coefs.

ProfileDP

The derivative of ProfileErr with respect to allpars[active].

ProfileList

Returns the results of ProfileErr and ProfileDP as a list with elements value and gradient

See Also

outeropt, Profile.LS, Profile.multinorm, LS.setup, multinorm.setup


[Package CollocInfer version 1.0.4 Index]