tracing optimization

[LAUJJL]

Hi everybody

I have two questions.
1. Is it possible when using the optimizer to have a report that traces only the different best pay offs, with the number of the simulation and the new pay off?
I need this because some of my optimizations, taking a long time , if I have to check the *.rep
Or *.out file, to find where a new best pay off comes, it may take a very long time.
I need to trace the time where new payoffs are generated because it will tell if the optimization process is likely to be finished after a definite time.
For instance, If I run a simulation for 12 hours, and I had 100 new payoffs the first 6 hours, and no one in the last 6 hours, it is probably because there is no more better payoff to be found.
It is in fact difficult to control the level of convergence during an optimization process.
I think that this method is maybe useful.
Generally I avoid using if then else in my models, to ease the Powell optimizing process.

2. One of my model has no dynamic at all. It has no levels in it, and I need to run it only for 1 step. I could use traditional tools to do that like Lingo, MPL or OPL Studio , but I prefer to use Vensim, because you can use it like an influence diagram, and express constraints (You put them in the payoff) easily. It is too easier to use then a spreadsheet and a solver because you can use subscripts.
To run it for only one step, I set the final time = initial time, the time step = 0, and when run the model, indicates an error, but after having calculated the only step, which is enough for me. But the problem comes with optimization. When run the model still indicates an error after the first simulation and then stops. The solution is then to set the final time = initial time + 1 and the time step = 1. It optimizes correctly, and gives the same results for both periods.
My model has about 80 parameters to be optimized. (It is difficult to reduce the number of
parameters, because they are highly independent, non dominant, while sharing common resources).

If I could optimize only on one period it would halve the time necessary to do it. Is it possible?
The model is deterministic (it has no stochastic data).
Regards.
J.J. Laublé

[bob@vensim.com]

To See when there are improvements in the payoff just put output level to 2 and then look at the output window that appears - it is updated everytime the payoff improves - this gets to runname.log as well. It is not, unfortunately, possible to scroll this window as the optimization progresses.

For a model with no dynamics run it from 0 to 1 with time step of 1. This will make twice as many computations as you need (actually more than twice because of initilization) but should still be fast relative to the alternatives you mentioned.

Make sure you are compiling the model if optimization time is an issue.

Bob Eberlein

[Edited on 26-5-2005 by bob@vensim.com]

[LAUJJL]

Thank’s Bob

What is the average percentage of improvement when compiling and is there a substantial increase of speed to prefer one compiler to another?
Thanks in advance and best regards.
J.J. Laublé.

[bob@vensim.com]

Hi JJ

It is usually 2 to 3 times as fast. The more subscripts, the bigger the improvement.

Microsoft has made their compile engine available without charge - go to http://msdn.microsoft.com/visualc/vctoo ... fault.aspx

I think you can get this in France as well as the US.

Bob eberlein