Ventana software support forum

Posted by Erling Moxnes <Erling.Moxnes@ifi.uib.no>
Thanks to Krys, Alan and Jean-Jacques for replies to my question about
optimisation. We seem to agree that a danger with the optimisation technique
is that it may lead to overly simplified models. Alan points out that this
was the explicit, historical reason for Forrester to be sceptical of
optimisation. Alan also makes clear that SD is not a branch of optimisation.

However, none of the three replies really answer my question. I was not
asking about reasons not to use optimisation techniques. I asked for reasons
not to formulate problems as ""optimisation problems"", while performing
standard system dynamics studies.

I agree with Krys that stakeholders may have different views on the goal,
and with Krys and Jean-Jacques that in most situations there are multiple
criteria that should be considered. This imposes challenges for the analyst.
This challenge may be ""solved"" by focusing on one aspect only, for instance
the economist may focus on a simple measure of profits and the system
dynamicist may focus on stability. However, both analysts may get it only
half right. Thus, both may benefit from framing the problem as a
multicriteria optimisation problem, even if they end up doing what they
usually do. However, by using this frame, they may at least force themselves
to argue why stability or why profits do not matter in their respective
analyses. Or, even better, they may end up augmenting their studies to
include important aspects that would otherwise have been left out.

Going back to the original theme for this thread, ""evaluating expected
modeling benefits"", it seems to me that if the analyst considers all the
most important criteria for the stakeholders (using the mantra ""what should
ideally be 'optimised'""), the benefits of the analysis would become
apparent. The standard policy could be compared with a policy that comes out
of the model study, and stakeholders can compare the different criteria and
judge the benefits for themselves. Now, this is of course what many or most
system dynamicists do. My point is that we may become even better at this if
we frame our problem statements in terms of multicriteria optimisation.

My best,

Erling Moxnes


-- The System Dynamics Group University of Bergen, Norway http://www.ifi.uib.no/sd/
Posted by Erling Moxnes <Erling.Moxnes@ifi.uib.no>
posting date Thu, 21 Apr 2005 12:38:15 +0200

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

I think that there will never be a definitive answer to your question,
because it depends on the problem you are working on.

The tendency is always to generalize something that has worked on something
to everything else, instead of trying to understand why one solution worked
and could not have worked if things had been different.

I too did not get any answer to my original question about expected
profits . which was very practical and problem oriented, and that was
transformed into a theoretical question. There is probably too no real
answer to my question.

It is too dependant on what you mean by optimization.

Any problem solving tends always to optimize or at least better something.

The problem of optimizing is that it reduces the ability to invent something
radically different because you are obliged to define a goal and by doing
this you freeze partially the situation avoiding any deep modification.
Optimizing is looking for a perfection that sometimes does not exist. There
is not one unique way to build the future.

Two examples:

I have spent about 2000 hours of work on a pricing optimization model, and
have lately understood that the way I had formulated the problem (an
optimization with dynamics), there was no practical answer worth to be
considered. There was probably a theoretical answer but to make that
theoretical model real was not possible for me.

Any problem can in theory be solved with SD, as reality is made of a
multitude of stocks that vary during the time depending on what they get and
what they loose 'the rates'.

On the contrary I had once a problem that I solved in about 4 hours with the
optimization features of Vensim to my complete satisfaction. It is in fact
the only model I was satisfied with so far. It models a German company named
Accenta that proposes to its clients when they buy something, to give them
about 7 % of the value of the good and they will pay them back the entire
value of the good plus the 7% the following 10 years of course with an
increasing yearly payback.

It is obviously a bubble. It is legal in Germany as long as there is a small
but important remark that says that if the payment is impossible it will be
postponed until it will be completely paid.

The business has splendid results because their promises are not considered
as debts.

The problem I solved very simply was the necessary increase of their
turnover to be able to carry on paying back what they had promised. I found
about 48% a year.

It was then easy to find at what level of increase they were only able to
refund the 7% plus the interests, that was the condition for the customer
not to loose anything depending on any interest rate. Other calculations are
easy to do. If the rate of increase is of 20% for example, how much will
increase the refund duration? Etc.

One can then add, the effect of an eventual postponing of the refund on
their turnover that is depending on the confidence of their clients and
build in advance counteracting policies.

Comparing both problems, I have approximately understood why I could solve
quickly one completely, while I failed to solve the second. The first has
too many dependant and non dominant factors and the second is very simple
and can even be formulated to a certain extent with no SD at all.

Regards.
J.J. Laublé Allocar
Strasbourg France
Posted by =?iso-8859-1?Q?Jean-Jacques_Laubl=E9?= <jean-jacques.lauble@wanadoo.fr>
posting date Thu, 21 Apr 2005 16:58:33 +0200

Posted by George A Simpson <gsimpso4@csc.com>
There is an article in the April 2005 Scientific American that addresses
your point: ""Shaping the Future"" by Steven Popper, Robert Lempert, and
Steven Bankes. The point they make is that real decision-making is more
sophisticated than simple optimisation.

The question is changed from ""what set of policies gives the optimum return
on investment (or other outcome metric)?

It becomes ""what actions today will best shape the future to our liking"".

What decision-makers are seeking is ROBUST policies, not necessarily
""optimal"" ones. Therefore the role of SD modelling is to allow a large
number of scenarios to be explored, identifying ones that have a good
outcome at low risk.

The authors also argue that incremental policy decisions, allowing for the
uncovering of new evidence and new influences, is a more powerful strategy
than one-shot optimisation.

We're using SD to probe the threads of the multiverse, trying to find thick
threads that lead in the direction we want to go.

..george...

Dr. George Simpson, Principal Consultant, CSC
CSC Alliance: Performance Engineer
CSC House, Fleet
Posted by George A Simpson <gsimpso4@csc.com>
posting date Thu, 21 Apr 2005 13:20:03 +0100

Posted by Bill Harris <bill_harris@facilitatedsystems.com>
>> Now, this is of course what many or most
>> system dynamicists do. My point is that we may become even better at this if
>> we frame our problem statements in terms of multicriteria optimisation.


I've found the ideas behind soft systems (the multiple CATWOEs, the
""rich picture"") to help in breaking away from the single criterion
approach.

Bill
- --
Bill Harris http://facilitatedsystems.com/weblog/
Facilitated Systems Everett, WA 98208 USA
http://facilitatedsystems.com/ phone: +1 425 337-5541
Posted by Bill Harris <bill_harris@facilitatedsystems.com>
posting date Thu, 21 Apr 2005 09:03:19 -0700