Difference between fmincon 'TolFun' and 'FunctionTolerance'

Hi all,

Recently I moved from Matlab R2015b to R2016b. One of the functions I use is fmincon. According to the help files for R2015b and R2016b, I conclude that the following options should be equivalent:

optimoptions(...
  @fmincon,...
  'Algorithm',                 'interior-point',...
  'Display',                   'off',...
  'TolFun',                    1.0e-4,...
  'StepTolerance',             1.0e-4,...
  'ConstraintTolerance',       1.0e-2,...
  'SpecifyObjectiveGradient',  true,...
  'SpecifyConstraintGradient', true,...
  'HessianFcn',                "MyHessian");

or:

optimoptions(...
  @fmincon,...
  'Algorithm',                 'interior-point',...
  'Display',                   'off',...
  'FunctionTolerance',         1.0e-4,...
  'StepTolerance',             1.0e-4,...
  'ConstraintTolerance',       1.0e-2,...
  'SpecifyObjectiveGradient',  true,...
  'SpecifyConstraintGradient', true,...
  'HessianFcn',                "MyHessian");

However, when I simulate my program with 'FunctionTolerance' I obtain different results in comparison to 'TolFun'.

Can anyone explain why this is happening?

Kind regards, Laurent Keersmaekers

Answer 1

[Alan Weiss]

I think that you forgot to specify OptimalityTolerance = 1e-4. The former TolFun was split into two new tolerances. As you see in the Current and Legacy Option Name Tables, the old TolFun option had two meanings: change in function value, and size of first-order optimality measure. These meanings are now codified in FunctionTolerance and OptimalityTolerance. OK?

Alan Weiss

MATLAB mathematical toolbox documentation

[Laurent Keersmaekers]

Thank you very much for your advice, this has indeed solved the issue.

Kind regards,

Laurent Keersmaekers