Ventana software support forum

Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
Hi Francisco.



I have more or less the same problem.

It is like dynamic programming in O.R.

One has to take decision every time step.

This decision depends on the past events and on future events that are
stochastic.

So you have to choose a policy that takes into account the future events
plus the future policies that may be taken.



I use Vensim, and the solution is not really embedded in the software, but I
use expedients to solve it, that run much more quickly than with Excel using
a solver, which is for me absolutely impracticable, as I simulate about 100
parallel simulations to simulate the stochastic future events and together
with optimisation and numerous parameters would take days to perform.



I take a simple example but it works for any other one.

Suppose that you have the periods P1..P4 4 periods, S1..S4 the policies to
define for each period, and the constraints for every period represented by
Vi <= 0, i = 1,4 where vi is a variable of the model, auxiliary or level.
The objective is to maximize for every period an objective function Oi i =
1,4 Oi = Oi(Si)

Maximizing the function Oi is equivalent to maximizing the sum of (Oi(si))
on the 4 periods.

To represent the constraints every step of the simulation I define a
function ci(vi) = (if then else (vi <= 0,0,-1000000)).


So you have converted your optimization step by step into en overall
optimization solvable by Vensim. the only thing to do is to optimize the
function F = sum (Oi(si)) + sum (ci(vi)).


What are the optimization parameters? All the Si policies together.


It works well when you have not too much periods, not too much parameters
for the si, and non stochastic models. Otherwise, the solution to minimize
the number of parameters to optimize is to define optimization rules as
parameters, that will generate the policies for each period automatically.
This may reduce considerably the number of parameters to optimize.

For instance with the first method if you have 10 parameters to optimize
each period, this results for a 100 period model in 1000 parameters to
optimize. Generally optimizing with rules gives less accurate result then
optimizing directly the parameters, but is becomes unavoidable with middle
to big models.

Of course finding the right rules and the parameters of the rules is not
evident

In the first solution of direct optimization of the parameters, you can end
up with a no loop model, while the rules solution have loops.

It may too be necessary to use external functions written in a procedural
language, where complex algorithms are much more easy to write then in
Vensim or Powersim. In Excel you will have too to use VBA.

Regards.

J.J. Laublé Allocar

Strasbourg France.
Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
posting date Mon, 29 Aug 2005 14:54:27 +0200

Posted by Jim Duggan <jim.duggan@nuigalway.ie>
Hi,

The following paper is definitely worth a look (this
years conference) as it explores ways of combining
decision analysis and system dynamics.

Osgood, Nathaniel Combining System Dynamics and Decision Analysis for
Rapid Strategy Selection

http://www.systemdynamics.org/conf2005/indexpapers.htm

Also, I think it's useful to make a distinction between
(1) SD models that call complex decision logic and (2)
complex decision logic that invoke SD models. In the
first case, the decision logic uses information cues from the model to generate a value of decision variable, in the second case (e.g. optimisation), the SD model is called repeatedly by an algorithm in order to minimise or maximise a payoff function.

It sounds like your task falls under the first
category.

Also, if you want to invoke a compex decision on
an infrequent basis (from Vensim, Stella, Powersim, etc.,) the MODULO inbuilt function is very useful (using the current simulation time and the decision time interval), and of course the decision can be encapsulated in some form of compatible external function (using a programming language).

best regards,
Jim.
Posted by Jim Duggan <jim.duggan@nuigalway.ie>
posting date Wed, 31 Aug 2005 10:13:41 +0100