MCSRCH Error

[GoogleCodeExporter]

Upon running interpolate-ngram twice, I received the same error twice after
the 46th iteration. When I did this with 1 less component lm, it went
through 188 iterations without error. 

interpolate-ngram -i GLI -op dev.txt -wl GLI.lm  -if
"log:sumhist:%s.effcounts" -o 6  -l "1.lm, 2.lm, 3.lm, 4.lm, 5.lm"

Optimizing 9 parameters...

 IFLAG= -1
 LINE SEARCH FAILED. SEE DOCUMENTATION OF ROUTINE MCSRCH
 ERROR RETURN OF LINE SEARCH: INFO=  3
 POSSIBLE CAUSES: FUNCTION OR GRADIENT ARE INCORRECT
 OR INCORRECT TOLERANCES
Iterations    = 46
Elapsed Time  = 516.460000
Perplexity    = 28.741343

Original issue reported on code.google.com by peppe...@gmail.com on 27 Feb 2009 at 10:56

[GoogleCodeExporter]

You get this error when the numerical optimization technique runs into a 
problem.  In 
this case, because we are estimating the gradient using a first-order 
difference, 
there may be some numerical instabilities.  For now, I would suggest trying a 
different optimization technique using the -opt-alg argument.  I plan to modify 
the 
implementation of parameter reading/writing so that you can start the 
optimization 
with random parameter values.

I am currently moving from MIT to Microsoft Research.  Thus, I will not be 
responding 
to issues as quickly.

Original comment by bojune...@gmail.com on 3 Mar 2009 at 4:16

• Changed state: Accepted