outputfcnsanitycheck

 function stop = outputfcnsanitycheck(params,optimValues,state,tol,numsteps)

 <params> is the current optimization parameter
 <optimValues>,<state> are optimization state stuff
 <tol> (optional) is a positive number.  default: 1e-6.
 <numsteps> (optional) is the positive number of iterations in the 
   past to compare to.  default: 10.

 we look back <numsteps> iterations and check whether the resnorm field
 of <optimValues> has changed by less than <tol>.  if so, we stop the
 optimization.  the only reason for this is to patch up some weird cases
 where the regular optimizer doesn't stop when it should (in these cases,
 the optimizer goes on forever without anything actually changing).
 
 example:
 x = -1:.1:20;
 y = evaldoublegamma([1 1 1 1 2 .2 0 0],x);
 yn = y + 0.1*randn(size(y));
 [params,d,d,exitflag,output] = lsqcurvefit(@(a,b) evaldoublegamma([a 0 0],b),ones(1,6),x,yn,[],[],defaultoptimset({'OutputFcn' @outputfcnplot}));

0001 function stop = outputfcnsanitycheck(params,optimValues,state,tol,numsteps)
0002 
0003 % function stop = outputfcnsanitycheck(params,optimValues,state,tol,numsteps)
0004 %
0005 % <params> is the current optimization parameter
0006 % <optimValues>,<state> are optimization state stuff
0007 % <tol> (optional) is a positive number.  default: 1e-6.
0008 % <numsteps> (optional) is the positive number of iterations in the
0009 %   past to compare to.  default: 10.
0010 %
0011 % we look back <numsteps> iterations and check whether the resnorm field
0012 % of <optimValues> has changed by less than <tol>.  if so, we stop the
0013 % optimization.  the only reason for this is to patch up some weird cases
0014 % where the regular optimizer doesn't stop when it should (in these cases,
0015 % the optimizer goes on forever without anything actually changing).
0016 %
0017 % example:
0018 % x = -1:.1:20;
0019 % y = evaldoublegamma([1 1 1 1 2 .2 0 0],x);
0020 % yn = y + 0.1*randn(size(y));
0021 % [params,d,d,exitflag,output] = lsqcurvefit(@(a,b) evaldoublegamma([a 0 0],b),ones(1,6),x,yn,[],[],defaultoptimset({'OutputFcn' @outputfcnplot}));
0022 
0023 % input
0024 if ~exist('tol','var') || isempty(tol)
0025   tol = 1e-6;
0026 end
0027 if ~exist('numsteps','var') || isempty(numsteps)
0028   numsteps = 10;
0029 end
0030 
0031 % global stuff
0032 global OFSC_RES;
0033 
0034 % do it
0035 switch state
0036 case 'init'
0037   OFSC_RES = [];
0038 case {'iter' 'done'}
0039   OFSC_RES = [OFSC_RES optimValues.resnorm];
0040   if length(OFSC_RES) >= numsteps+1
0041     if abs(OFSC_RES(end)-OFSC_RES(end-numsteps)) < tol
0042       stop = 1;
0043       return;
0044     end
0045   end
0046 end
0047 
0048 % return
0049 stop = 0;
0050 
0051 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%
0052 
0053 % % for sanity checking, we also check whether optimValues.ratio
0054 % % is NaN; if it is, we stop the optimization.
0055 % %
0056 % %   if isfield(optimValues,'ratio') && isnan(optimValues.ratio) && optimValues.iteration > 5
0057 % %     stop = 1;
0058 % %     return;
0059 % %   end