NEWS

---

gremlin 1.1.0 (2024-11-04)

NEW

• added c++ (and R) code to use the Takahashi et al. algorithm for obtaining the partial inverse of the coefficient matrix

  • This switch means almost a 20-fold speedup (for c++ routines) when using analytical first derivatives (i.e., need to calculate "trace" terms)

  • Speedup comes from only calculating elements of inverse matrix (C-inverse) following the non-zero pattern determined for the Cholesky decomposition of the C matrix.

  • follows the SuiteSparse Matlab_Tools sparseinv by Tim Davis, but sparseinv works on LDL' factorization of C whereas I changed this to work on LL' factorization of C.

    • Takahashi, Fagan, & Chin. 1973. Formation of a sparse bus impedance matrix and its application to short circuit study. 8th PICA Conference Proceedings, Minneapolis, MN.

  • gremlinR() now uses far less RAM per iteration of the model (previously was forming the entire C-inverse)

  • add finite difference algorithm to obtain first derivatives of likelihood function

    • introduced a parameter (h) inside gremlinControl() to set the "difference" or amount to alter parameters to calculate change in log-likelihood.

  • created a REML function inside the c++ code (reml) to calculate log-likelihood

    • moved log-likelihood calculation out of main program and reduced amount of repeated code.
    • also facilitated finite difference functions (which are just repeated log-likelihood evaluations)

Minor Changes

• Removed error() in c++

  • now issues with matrix singularities etc. do not stop code without returning model so far

  • should now be possible to use update() to get "through" trouble spots

  • also allows for user interruptions to c++ code from terminal

  • Changed default parameterization so lambda transformation is no longer the default

  • Changed default convergence check criteria (cctol)

    • Models using previous values tended to only improve precision of estimates well beyond what was meaningful.

gremlin 1.0.1 (2020-06-25)

NEW

• deltaSE() function to calculate approximate standard errors for functions of (co)variance parameters (e.g., h², standard deviations of variances, or correlations)

  • this can take a formula for the function or a character expression

  • also allows for a list of formulas or character expressions e.g., calculate all variance components as proportions of total variance

  • Introduce Gcon and Rcon arguments to gremlin() for constraining parameters

    • enables parameters to be fixed or otherwise constrained
    • works in conjunction with the Gstart and Rstart arguments
    • For example in a simple sire model, we could restrain the sire variance =0.38.

grSf <- gremlin(WWG11 ~ sex,
	random= ~ sire,
	data = Mrode11,
	Gstart = list(matrix(0.38)),
	Gcon = list("F"),
	control = gremlinControl(lambda = FALSE))

• Similar to above change (Gcon/Rcon), introduced steps to deal with parameters outside of the boundaries of their parameter space (e.g., variance < 0).

  • restrain these parameters to near their boundaries (after trying step-reduction calculation)
  • re-calculate Average Information, conditional on restrained parameters
    • See Gilmour. 2019. J. Anim. Breed. Genet. for specifics

• change version numbering to just 3 numbers (instead of 4)

  • just dropping last number

Minor Changes

• create new c++ function to handle quasi Newton-Rhapson algorithm
  • allows secondary checks of appropriateness/naughtiness for proposed parameters based on a conditional AI algorithm (conditional on parameters restrained to boundary condition)

gremlin 1.0.0.1

NEW

• update() function

  • can now continue a model where it left off or change the structure (e.g., drop a single variance component for likelihood ratio test)

  • Implement "step-halving" algorithm for AI updates

    • restricts parameter updates if AI algorithm proposes a change of >80% of original parameter value
    • amount by which a parameter change is restricted can be set in gremlinControl() using the step argument

Minor Changes

• Implement more efficient algorithms in the c++ code, that were developed in the R code for version 1.0.0.0.
  • Add gremlinControl() function for advanced changes to the way gremlin runs
  • Begin major improvements to speed of gradient calculation function
    • changes to be incorporated in em, ai, and elsewhere (where relevant) in next version
    • implements calculations that take advantage of sparsity (i.e., don't calculate values where there are zeroes)

gremlin 1.0.0.0

NEW

• Completely revised way models are built and called

  • made a "modular" series of functions for setting up the model and optimizing the REML likelihood

  • new grMod and gremlinR classes.

    • grMod is the model structure for which a log-likelihood can be calculated
    • gremlinR class distinguishes from gremlin class in that gremlinR objects will only use R code written by the package in order to run the model. Class gremlin will execute underlying c++ code written in the package.

  • Average Information algorithm has been vastly improved

    • ai() efficiently calculates the AI matrix without directly computing several matrix inverses (as previously coded)

  • lambda and alternative parameterizations now possible and executed by the same code

    • lambda parameterization is the REML likelihood of the variance ratios after factoring out a residual variance from the Mixed Model Equations.
    • the alternative does not have a special name, this is just a model of all (co)variance parameters as (co)variance parameters (as opposed to ratios, as in the lambda models).
    • instead of completely separate functions for these two parameterizations, there is an argument that runs alternative lines of code, wherever the calculations differ for these two different parameterizations

Minor Changes

• No long construct Mixed Model Array (M) matrix from which the Cholesky factorization (and logDetC and tyPy calculations are made)
  • Changed to directly construct coefficient matrix of mixed model equations (C) and obtain tyPy and logDetC using this
  • Previously had to store Cholesky factorizations of both M and C, now do a solve with Cholesky of C (sLc/Lc in R/c++ code) to calculate tyPy based off Boldman and Van Vleck

gremlin 0.1.0.0 (2018-10-30)

NEW

• methods for gremlin objects
  • notably, AIC, residuals, anova, and nobs
  • updated the summary, print, and logLik methods as well

gremlin 0.0.2.0

Improved algorithm that reduces computational resources and time! Also implemented c++ code in gremlin(), while keeping gremlinR() purely the R implementation (at least from the package writing standpoint).

gremlin 0.0.1.0

NEW

Documentation has switched from filling out the .Rd files manually to providing documentation next to the function code in the .R files using roxygen2

gremlin 0.0.0.1

Congratulations, its a gremlin!