Calibration using GA and then ANN

I have to do a calibration of a model with 10 parameters. Used GA to calibrate, but to avoid convergence of GA at local optima I wanna use ANN to find the global optimum. Can I use the s开发者_运维百科ame values of parameters obtained during GA for ANN. And How to do it ANN? My model is like Output = F (Input, Parameters). Thank you.

GA and ANN are equally likely to get stuck in local minima. Pick one and use one of the many techniques designed to avoid premature convergence.

With GA the most important would be to lower selection pressure, which will let you avoid local minima at the cost of slower convergence. Assuming you have time to wait make the selection pressure extremely low, so that better solutions are just slightly more likely to get picked in favour of worse. This is generally true for all heuristic parameter optimization - most people new to it will think quick, greedy convergence is a good thing. It is not. Try to make your convergence as slow as possible, which will let your algorithm explore as many weird alternatives as possible.

Other things that will help: increase population size, add demes (isolated populations that only "communicate" intermittently), rerandomize and run your optimization several times, add fitness penalties for being similar to the general population.

ANNs are not generally a good match for parameter optimization, as ANNs excel at pattern recognition, not exploring a problem space. Since you don't seem to have gone far down the ANN path, I would follow @Gurgeh's advice re. techniques to avoid local minima in your GA.

It sounds like a GA problem. ANN can't help. Generally speaking GA are used for optimization problems while ANNs are suitable for solving classification problems. GA can be used for ANN optimization. The weak point is the use of an ANN for calibration. Unless you can express the calibration result as a classification assumption, you should settle to GA using it's mechanisms and technics to avoid, as possible, local optima.