How to change the step size for fmincon?

Hello! I am trying to use fmincon to optimize some weights for an MPC controller. The weights can vary from about 0 to 1000 but the weights only change by about 0.001 each iteration. I would like it to change by whole numbers. I have tried to adjust FiniteDifferenceStepSize and DiffMinChange but have had no luck yet. Does anyone have experience with this? Thank you!

Nathan

[Mario Malic]

How many variables do you have and how many iterations there were before you stopped the solver?

If you have 10 variablles, function has to be evaluated ~10 times with variables changing for FiniteDifferenceStepSize value before solver starts moving with "bigger steps", Try patternsearch if you have it available.

Otherwise, RelLineSrchBnd and RelLineSrchBndDuration, but these are not easy to figure out how to choose and I can't really help with these, as they are hidden for a purpose.

[Nathan Batta]

I have 12 variables but the solver never gets to that many iterations. Usually around 5. And it never changes the variable by any noticeable amount even though FiniteDifferenceStepSize is very large

[Mario Malic]

FiniteDiffStepSize does not alter the step size between iterations. It affects only optimisation variables in between iterations (function evaluations). You should've included more information about the variable values, options, lb, ub, etc. Try setting RelLineSrchBnd to 1 or 5 or something like that. About duration, set it to 500 or something. After this number, IIRC, step size goes back to default value(very small).

[Nathan Batta]

Still no luck at getting it to optimize. The initial condition and bounds are seen below:

weights0=100*ones(1,12);
ub=1000*ones(1,12);
lb=zeros(1,12);

A typical output is:

99.775  100.381  101.0180  100.122  100.383  101.939  98.432  100.700  99.252  100.711  99.921  105.075

I am looking to minimize absorbed power but the output from fmincon is clearly not the minimum becuase I can call the function with other values and get lower values for power. Any input helps! Thanks

[Mario Malic]

Remember that fmincon is a local search option. In this case, where you don't know how your objective function looks like, it's just a matter of pure luck getting a very good solution.

weights0=ones(1,12); % no need to multiply it by a factor

Please read this, did you try with RelLineSrchBnd and RelLineSrchBndDuration?

[Nathan Batta]

I did try RelLineSrchBnd and RelLineSrchBndDuration but had little luck. The best results I've found is by making the OptimalityTolerance and StepTolerance extremely small (1e-80). However, it is still only varying by about 30 when I want it to step by much more.

[Mario Malic]

Well, fmincon searches locally. Plot your fval values, do they converge, if so, then local minimum is found. This could be a good problem for ga if your function evaluations are short.