[daip] Type of self-calibration solver used in CALIB

[Eric Greisen]

Roger Deane wrote:
> Dear Eric
> 
> I was hoping you could tell me what kind of solver is implemented during 
> self-calibration within classic AIPS? Is it the same as AIPS++ (i.e. the 
> Levenberg–Marquardt algorithm)?
> 
> We are considering implementing a Bayesian solver within a new package 
> called MeqTrees but this would require changing the current solver.
> 
> Do you know of any other efforts to do this?

There are several solvers, all quite unsophisticated, in AIPS.  They are 
in the
$APLNOT area (31DEC09/APL/SUB/NOTST), called from CLBPA and called 
GCALC, GCALC1, GCALC2, NCALC, and ECALC.  The help file for CALIB will 
list some general description of them - all offer a single pass or a 
"robust" multi-pass in the solver.   We have found no need in AIPS to 
implement anything further - in keeping with our normal philosophy, 
these are simple routines that can be managed and debugged.  There has 
been no need for something more elaborate and our experience is that 
elaborate is bad unless absolutely required.  AIPS used to report as 
failed solutions that failed because they were flagged in some IFs but 
not all.  This has been corrected and perhaps has led some to think that 
the solver had serious problems.  If the data are any good, there is no 
problem.  Perhaps, if there were e.g. 4000 antennas to solve, then a 
fancier solver might be required.

Eric Greisen