Differences

This shows you the differences between two versions of the page.

--- readme.aireml [2014/09/05 16:16] – shogo
+++ readme.aireml [2014/11/25 11:32] – shogo
@@ Line 2: / Line 2: @@
 ===== Summary =====
-A modification of REMLF90 for estimating variances with the Average-Information algorithm. Initially written by Shogo Tsuruta in 03/99-07/99.
+A modification of REMLF90 for estimating variances with the Average-Information algorithm. Initially written by Shogo Tsuruta in 03/99-07/99. AIREMLF90 uses a second derivative REML algorithm with extra heuristics, as is described in Jensen et al. (1996-7). For most models, it converges in far fewer rounds than EM-REML as implemented in REMLF90. While typically REMLF90 takes 50-300 rounds to converge, AIREMLF90 converges in 5-15 rounds and to a higher accuracy. The final results will be saved in "airemlf90.log".
 \\
-AIREMLF90 uses a second derivative REML algrithm with extra heuristics, as is described in Jensen et al. (1996-7). For most models, it converges in far fewer rounds than EM-REML as implemented in REMLF90. While typically REMLF90 takes 50-300 rounds to converge, AIREMLF90 converges in 5-15 rounds and to a higher accuracy. For selected models, AI-REML may fail to converge when the covariance matrix is close to non-positive definite. Adjust sensitivity of the program by setting the appropriate tolerance or setting good starting values. The final results will be saved in "AIREMLF90.log".
 \\
 See PREGSF90 with genotypes (SNP) for options.
@@ Line 128: / Line 127: @@
 <file>OPTION out_se_covar_function</file>
 Indicate to store in file samples of (co)variances function for postprocessing (histogram, etc.)
+===== Tricks =====
+When the covariance matrix is close to non-positive definite, the AIREMLF90 may not converge.
+There are two options you might want to try:
+. change the tolerance value (xx) in the option:
+OPTION tol xx
+to a very strict value (e.g., 1d-20) or a lenient value (1d-06).
+. use an option to use EM-REML inside AI-REML:
+OPTION EM-REML xx
+where xx is the number of iterations for EM-REML you expect to get a good starting value for AI-REML. After running xx rounds with EM-REML, the AIREMLF90 program will automatically switch from EM-REML to AI-REML using the last estimate from EM-REML as a starting value for AI-REML.