On 9/16/2016 11:40 AM, Hazhir wrote:
> I am running an optimization using fmincon interior-point algorithm
> (~100 parameters; scaled so all are in the 0-1 range). The algorithm
> converges due to exceeding TolX (I can go as small as 1e-12 and it
> converges), despite the fact that first order optimality measure is
> large and positive (~1e7) and graphing the function around the solution
> shows clear room for improvement. Why is then the step size going to
> zero despite the large gradient calculated by Matlab? What can I do to
> fix this problem? As a result, comparing with modified powell conjugate
> search implemented in Vensim simulation software, Matlab optimization
> can't find the local optima despite many different start points.
My generic advice is in
http://www.mathworks.com/help/optim/ug/when-the-solver-might-have-succeeded.html#br44i8w
or perhaps in
http://www.mathworks.com/help/optim/ug/when-the-solver-fails.html
Sorry, I cannot give more specific advice without more details, such as
your constraint types and some more information about your objective
function such as whether it is given by a simulation or ODE solution,
and does it have points of nondifferentiability.
Sorry you are frustrated,
Alan Weiss
MATLAB mathematical toolbox documentation
> I am running an optimization using fmincon interior-point algorithm
> (~100 parameters; scaled so all are in the 0-1 range). The algorithm
> converges due to exceeding TolX (I can go as small as 1e-12 and it
> converges), despite the fact that first order optimality measure is
> large and positive (~1e7) and graphing the function around the solution
> shows clear room for improvement. Why is then the step size going to
> zero despite the large gradient calculated by Matlab? What can I do to
> fix this problem? As a result, comparing with modified powell conjugate
> search implemented in Vensim simulation software, Matlab optimization
> can't find the local optima despite many different start points.
My generic advice is in
http://www.mathworks.com/help/optim/ug/when-the-solver-might-have-succeeded.html#br44i8w
or perhaps in
http://www.mathworks.com/help/optim/ug/when-the-solver-fails.html
Sorry, I cannot give more specific advice without more details, such as
your constraint types and some more information about your objective
function such as whether it is given by a simulation or ODE solution,
and does it have points of nondifferentiability.
Sorry you are frustrated,
Alan Weiss
MATLAB mathematical toolbox documentation