[MITgcm-devel] restarting an optimization

[Martin Losch]

Hi Patrick (and others),

it was easy to manipulate the_model_main to pack the ctrl vector again at the end of the iteration: just modify/remove the corresponding if statements around the second call of ctrl_pack.

The second thing I had to do, was to "cold start" the optimizer (remove OPWARMI and OPWARMD), only this removed the unwanted features in the resulting xx_* files of the next iterations. If I understand the minimization algorithm BFGS correctly, OPWARMD contains information from previous iterations (mainly gradient information in the form of an estimate of the inverted hessian). In my case this gradient used to be wrong a few iterations ago, and this was apparently still reflected in OPWARMD and consequently in the descent direction lsopt chose.

Martin


On May 10, 2011, at 6:21 PM, Patrick Heimbach wrote:

> Hi Martin,
> 
> the control vector is only packed at iter. 0
> (since no packed file exists then).
> For iter. > 0 only the gradient is packed, i.e. ecco_cost...$iter is generated
> (since ecco_ctrl...$iter is already available as the output from the line search).
> Should be easy to "trick" the model to pack both
> (in S/R the_model_main).
> 
> -p.
> 
> On May 10, 2011, at 11:59 AM, Martin Losch wrote:
> 
>> Hi there,
>> 
>> rather than repeat the entire optimization, I'd like to restart one at some iteration, after resetting one of the control parameters. In my case, I want to modify xx_theta (initial conditions for theta). I thought that when I modifiy xx_theta.${iteration}.data while the adjoint model is running, ctrl_pack should use the modified version and pack it into the control vector that is read by optim.x
>> However, the result is discouraging: nothing substantial changed. What else do I need to reset such an iteration? Force a cold start by the optimizer?
>> 
>> Martin
>> 
>> 
>> _______________________________________________
>> MITgcm-devel mailing list
>> MITgcm-devel at mitgcm.org
>> http://mitgcm.org/mailman/listinfo/mitgcm-devel
> 
> ---
> Patrick Heimbach | heimbach at mit.edu | http://www.mit.edu/~heimbach
> MIT | EAPS 54-1518 | 77 Massachusetts Ave | Cambridge MA 02139 USA
> FON +1-617-253-5259 | FAX +1-617-253-4464 | SKYPE patrick.heimbach
> 
> 
> 
> _______________________________________________
> MITgcm-devel mailing list
> MITgcm-devel at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-devel