Repeated fmincon optimization of slightly different objective functions

Question

Rosi Marungu 2016-5-6

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/282910-repeated-fmincon-optimization-of-slightly-different-objective-functions

评论： Rosi Marungu 2016-5-6

Hi!

I have a (log-likelihood) function f(x,data), which I optimize repeatedly w.r.t x using fmincon. After each optimization one datapoint is added and I use the previous solution as the starting value for the next optimization. There are many datapoints so the optimization problem and solution only change marginally. However the optimizer diverges each time only to converge back to a solution close to the starting point. This is very time consuming. I cannot provide an analytical gradient or hessian.

How could I speed up the procedure? I thought about passing in an initial gradient and hessian, but all I could find is an option to pass in an analytical gradient and hessian.

Thank you

Rosi

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Alan Weiss 2016-5-6

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/282910-repeated-fmincon-optimization-of-slightly-different-objective-functions#answer_221080

I am not sure why you are using the procedure that you describe. Do you need the intermediate solutions after each set of data points, or are you simply trying to guide fmincon to the global solution?

If you are simply guiding fmincon, then I suggest that you try giving fmincon all the data points at the start, and instead start fmincon from a variety of initial points to search for a global solution.

If you need the intermediate solutions, then I do not have any ideas. Sorry.

Alan Weiss

MATLAB mathematical toolbox documentation

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Rosi Marungu 2016-5-6

Thank you for your answer.

Unfortunately, I need the intermediate solutions. I try to evaluate the predictive power of my time series model. At each t I need the optimization solution to predict t+1. Then I compare all forecasts to the real data.

请先登录，再进行评论。