aboutsummaryrefslogtreecommitdiff
path: root/NEWS.md
diff options
context:
space:
mode:
Diffstat (limited to 'NEWS.md')
-rw-r--r--NEWS.md16
1 files changed, 8 insertions, 8 deletions
diff --git a/NEWS.md b/NEWS.md
index 0a2b3300..37964fd9 100644
--- a/NEWS.md
+++ b/NEWS.md
@@ -1,6 +1,6 @@
-# mkin 0.9.49.5 (2019-06-04)
+# mkin 0.9.49.5 (2019-07-03)
-- Several algorithms for minimization of the negative log-likelihood for non-constant error models (two-component and variance by variable). In the case the error model is constant variance, least squares is used as this is more stable. The default algorithm tries direct minimization and a three step procedure, and returns the model with the highest likelihood.
+- Several algorithms for minimization of the negative log-likelihood for non-constant error models (two-component and variance by variable). In the case the error model is constant variance, least squares is used as this is more stable. The default algorithm 'd_3' tries direct minimization and a three-step procedure, and returns the model with the highest likelihood.
- The argument 'reweight.method' to mkinfit and mmkin is now obsolete, use 'error_model' and 'error_model_algorithm' instead
@@ -8,19 +8,19 @@
- New function 'mkinerrplot'. This function is also used for residual plots in 'plot.mmkin' if the argument 'resplot = "errmod"' is given, and in 'plot.mkinfit' if 'show_errplot' is set to TRUE.
-- Remove dependency on FME, only use nlminb for optimisation
+- Remove dependency on FME, only use nlminb for optimisation ('Port' algorithm). I cannot remember cases where one of the other optimisation algorithms was preferable, except that I sometime used Levenberg-Marquardt for speed in cases where I did not expect to get trapped in a local minimum.
-- Use the numDeriv package to calculate hessians
+- Use the numDeriv package to calculate hessians. This results in slightly different confidence intervals, takes a bit longer, but is apparently more robust
-- Add a benchmark vignette to document the impact on performance. For very simple fits, the new code is a bit slower, presumably because of the time it takes to calculate the hessian matrices with and without parameter transformation
+- Add a simple benchmark vignette to document the impact on performance.
-- The code for manual weighting was removed
+- The code for manual weighting was removed. This functionality might get added again at a later time.
-- The fitting time reported in the summary now includes the calculation of the hessians
+- The fitting time reported in the summary now includes the time used for calculation of the hessians
- Adapt tests
-- Fix an error in the FOCUS chi2 error level calculations that occurred if parameters were specified in parms.ini that were not in the model. A warning was already issued, but when fitting via mmkin this could easily go unnoticed.
+- Fix an error in the FOCUS chi2 error level calculations that occurred if parameters were specified in parms.ini that were not in the model. A warning was already issued, but when fitting in parallel via mmkin this could go unnoticed.
# mkin 0.9.48.1 (2019-03-04)

Contact - Imprint