I think when using 600 variables, the solver does not enter the newton iterations (I stop it before).
My problem is a least squares problem so using a quasi-newton algorithm (which only requires to compute the derivative of the model to have good estimates of the derivative and Hessian of the cost function) would probably be more adapted (quicker convergence). My cost function to minimize has some local optimum but I can enter a first guess close enough to the global minimum to not get stuck in a local minimum. Next step would be to use a simulated annealing algorithm to be free of choosing the initial guess. Best regards, Frédéric Parrenin 2013/2/4 Morten Welinder <[email protected]> > Ouch. 600 is well outside the design range. > > You might have to play with neutering parts of the non-linear > solver. For example, basically turning off the newton steps > because computing the Hessian means something like > 2*600*600 recalculations of your target cell. > > Also, the polishing steps are probably a waste of time for you. > > See also http://www.solver.com/hard-models -- that is written > about a different solver but the basic points apply here too. > Notably > > > If the problem is convex, you can have very high confidence > > that the solutions obtained are globally optimal. If the problem > > is non-convex, you can have reasonable confidence that the > > solutions obtained are locally optimal, but not necessarily > > globally optimal. > > M. >
_______________________________________________ gnumeric-list mailing list [email protected] https://mail.gnome.org/mailman/listinfo/gnumeric-list
