consequences of optimising all subsystems

[Erling Moxnes]

Economics deal with "optimising all subsystems" in its literature on the
"second best". In general, when one policy is not "first best", it is not
obvious that one should strive for fist best policies to other problems.

See for instance:
Copeland, B.R. (1994). "International Trade and the Environment: Policy
Reform in a Polluted Small Open Economy." Journal of Environmental Economics
and Management 26(1):44-65.

Best regards, Erling Moxnes
From: Erling Moxnes <Erling.Moxnes@snf.no>

-----------------------------------------------------------------------
Erling Moxnes
SNF, Breiviken 2, 5035 Bergen-Sandviken, Norway
Tlf.+47 55 959526
Fax.+47 55 959439
http://www2.snf.no/~erling/

["Paul J. Campbell"]

Braesss Paradox: Opening a new highway, or increasing the capacity of an
existing street, can lengthen average trip times.

See article "Braesss Paradox: A Puzzler from Applied Network Analysis"
by Steve McKelvey in The UMAP Journal 13 (4) (1992) 303-312.

Authors address: mckelvey@stolaf.edu



THROUGH AUGUST, 1998
Prof. Paul J. Campbell
c/o Lst. Prof. Pukelsheim
Institut fuer Mathematik der Universitaet Augsburg
Universitaetstrasse 14
D-86135 Augsburg
Germany

+49-821-598-2162 (ofc)
+49-821-598-2280 (fax)
+49-821-811-986 (res)
(from U.S., substitute 011 for +)
Central European Time: 7 hrs later than Chicago

campbell@math.uni-augsburg.de
http://cs.beloit.edu/campbell/ (www pages)


PERMANENT ADDRESS, AND FROM 9/98:
Paul J. Campbell
Mathematics and Computer Science
Beloit College
700 College St.
Beloit, WI 53511
(608) 363-2007 (ofc)
362-2805 (res)
363-2718 (fax)
campbell@beloit.edu (email)
http://cs.beloit.edu/campbell/ (www pages)

[Jerome Winston]

*Preface: This may not be the most appropriate list on which to ask raise
this topic, however it is the only systems list of which I am aware. If
there are other lists on which it would be more appropriate for me to raise
this question, please let me know. Thanks.*

My question relates to the consequences of attempting to optimise
subsystems. I have a vague memory of reading about this matter, way back
in the early 1970s, perhaps in the Bulletin of Atomic Scientists (perhaps
elsewhere) -- in a discussion of the (ir)relevance of criticisms of The
Club of Rome Report.

What I recall is a reference to a theorem that asserts that under certain
general conditions, optimising all the (interconnected) sub systems in a
sytem will lead to (a) an unstable total system that (b) is itself not
optimised.

My question is: Is there such a theorem (or a similar theorm), and if there
is, where can I locate a proof and/or discussion of it?

Comments: OR textbooks and teachers give examples of how optimising a
total system can occur even though one or more sub systems are not
optimised. In the books that I have seen, such examples are cited as
special cases, not as examples of a general theorem.

My interest in the relationship between optimising of subsystems and
optimisation of a total system arises from my work in public sector
evaluation.

Around the world, consultants are advising governments that it is important
to optimise the "performance" of every sub system. This is often given as
one of the main reasons for "performance measurement": to allow the
performance of every sub system to be optimised.

To what extent can I assert that *theory* provides a basis for rejecting or
modifying this advice?

I have practical examples that appear to support the conclusion that it is
unwise to proceed without exception to optimise all sub systems, but I can
only call on "common sense" to support my conclusion. I recognise the
fallibility of ill-informed "common sense" when dealing with systems!

Hence, before I stick my neck out too far, Id like to establish the extent
to which I can cite a theoretical basis for asserting that (under certain
conditions) it is unwise to seek to optimise a total system by optimising
all subsystems.

I would appreciate any help list members can provide me (via the list or by
direct email) to (a) (re)formulate my questions so they can be answered
and/or (b) track down whatever theory is available that sheds light on the
consequences of optimising sub systems. Thanks.

Jerry Winston
From: Jerome Winston <jwinston@rmit.edu.au>

--------------------------------------------------------------
Jerome A. Winston,
Director, Program for Public Sector Evaluation (PPSE)
Royal Melbourne Institute of Technology
>From overseas: Office Phone 61-3-9468-2387; Fax 61-3-9467-8708

[Stephen Shervais]

Jerome,
There is some recent work on self-organizing systems that gives insight
as to when this might or might not work. Stuart Kauffman has written
about it in two books:

"The Origins of Order" Oxford University Press 1993 ISBN 0-19-507951-5
"At Home in the Universe" OUP 1993 ISBN 0-19-511130-3

The first is the more techical, and the information is to be found in his
discussion of "homogeneity clusters" in random boolean nets. The second
is a little more accessible, and talks about "patches."

The basic idea is that there is some natural size to the sub-optimized
patch that will allow the entire organization to achieve some energy
minimum. If your patches are too large -- the organization is too
centralized (what he calls "Stalinist") -- you freeze. If the patch is
too small -- ("Leftist Italian") -- you get chaos. The best place to be
_seems_ to be with a patch size that keeps you right at the edge of chaos.

I have not seen any SD references to this, and am not sure how SD plays
in the random boolean arena. Perhaps others could extend and modify my
remarks.


Regards,
Steve

Steve Shervais shervais@acm.org
Graduate Student shervais@sysc.pdx.edu
Systems Science PhD Program psu00872@odin.cc.pdx.edu
Portland State University http://www.sysc.pdx.edu
Portland, Oregon http://www.ee.pdx.edu/~shervais
(503) 725-7344 / 725-4997

"You think because you understand ONE
you must understand TWO,
because ONE and ONE are TWO.
But you must also understand AND."

- Sufi wisdom via Meadows
through Wheatly

[George Richardson]

I recall hearing one of the authors of "Mankind at the Turning Point" by
Mesarovic and Pestel (a book published by the Club of Rome after "Limits
to Growth") claim the same theorem mentioned in this query. Mr. Winston
may find a reference in that book to the theorem and its source(s).

Ive always been curious about this myself, so I hope to hear a definitive
answer to this query.

...GPR

-----------------------------------------------------------------------
George P. Richardson G.P.Richardson@Albany.edu
Rockefeller College of Public Affairs and Policy Phone: 518-442-3859
University at Albany - SUNY, Albany, NY 12222 Fax: 518-442-3398
-----------------------------------------------------------------------

["Pelham Barton"]

Jerry Winston said:

>Around the world, consultants are advising governments that it is
>important to optimise the "performance" of every sub system. This is
>often given as one of the main reasons for "performance measurement":
>to allow the performance of every sub system to be optimised.

An immediate point that occurs is that one should try to insist that
performance measures are defined in terms of the system as a whole.

I used to be a schoolteacher. The usual performance measure for
"satisfactory discipline" is that children feel unable to misbehave
in your lessons. What if this is achieved by building up resentment
in the children, which is then taken out either on your colleagues,
or on weaker pupils through bullying?

I am sure that similar examples can be found in other areas.

Pelham Barton
University of Birmingham
p.m.barton@bham.ac.uk

[hamersma.maarten@columbus.co.za]

GDay Everyone
Jerry had a question re. optimising subsystems.
snip>>
>
>What I recall is a reference to a theorem that asserts that under certain
>general conditions, optimising all the (interconnected) sub systems in a
>sytem will lead to (a) an unstable total system that (b) is itself not
>optimised.
>
>My question is: Is there such a theorem (or a similar theorm), and if there
>is, where can I locate a proof and/or discussion of it?
>
snip>>

Have a look at the "Theory of Constraints", an "invention of
Eli Goldratt of the Goldratt Institure - he developed a
production planning/scheduling theory based on global
optimisations, and on the way discussed various issues
regarding "local optimums", or optimised subsystems.

He wrote a few books: The Goal, The Haystack Syndrome and others.
In Aus. there is a company, "Scheduling Technology Group", I think they
are called, which will have details. Otherwise hit the web, Im
sure amazon.com (Amazon Books) will have references.
You could also talk to Gene Bellinger <gbellinger@outsights.com>,
who knows a lot about these things.

Regards,

Maarten Hamersma Phone: +27 (0)13 247 2235
Systems Engineer Fax: +27 (0)13 246 1108
Columbus Stainless, South Africa
From: hamersma.maarten@columbus.co.za (Maarten Hamersma)

[ularoch@ibm.net]

even in nonlinear control theory it is concluded, that optimizing subsystems
only is a sure way of not reaching the overall optimization possible. beeing
a zynic i conclude, that a "good" consultant tries to have always some work
left over....hence only advising work with subsystems.
using discrete-event tools will be a sure way not to tackle optimizing
systems top-down, but also SD could of course be misused to that purpose.

// yours sincerely ulrich la roche
fast focus consulting
heilighuesli 18, CH-8053 Zuerich,
switzerland
fax +411 382 1349
From: ularoch@ibm.net

[Alex Rodrigues]

"What I recall is a reference to a theorem that asserts that
under
> certain general conditions, optimising all the (interconnected) sub
> systems in a system will lead to (a) an unstable total system that (b)
> is itself not
> optimised. My question is: Is there such a theorem (or a similar
> theorem), and if there is, where can I locate a proof and/or
> discussion of it?"
>
Jerome, let me have a simple but straight-forward thought about
your very interesting query.

The relationship between the individual optimums of each
sub-system and the global system optimum relies on two main factors: (1)
the formulation of the global optimum from the individual performances
of each sub-system, and (2) the competitive or symbiotic interactions
among the sub-systems, while striving to reach their optimums (which may
restrict or reinforce each others performances).

As a simple example of (1) consider a system with two
sub-systems A and B, and P(A) and P(B) being the performances of each,
respectively. Assume no interactions between the two systems. Now, if
the performance of the whole system is formulated as, P(S)=P(A)+P(B),
then it is obvious that the maximum performance of the global system is
achieved when its sub-systems are maximised. Now, consider a formulation
like:
P(S) = 4*P(A)*P(B)-P(A)^2-P(B)^2. Then, most of us know for sure
that the maximum of the system is NOT achieved when both sub-systems try
to achieve their maximums (e.g P(S)[A=3, B=2]=11 > P(S)[A=7, B=2]=3).

Regarding (2), it is obvious that when the sub-systems compete
for maximum performance, then the marginal gains and losses will
influence whether it is better to sacrifice performance of A in favour
of B (or vice-versa), and in what conditions. If the interaction is
symbiotic then improving one system will help to improve the other.
However, this might also be at the expense of sacrifying the performance
of other sub-systems; the final impact on overall system performance
will again also depend on (1).

Although the above examples refer to illustrative mathematical
formulations, far simpler than reality, the message is that you may want
shift your search for a universal theorem -- basically saying that (*)
"The whole is more than just the sum of its parts" -- into a theorem or
research that describes or classifies the performance of real systems in
terms of (1) and (2). With this information at hand, you may then
deduce whether (*) applies or not, and in what conditions.

I find it hard to believe that for all systems, (1) and (2) are
such that an universal theorem, saying that you must not try to optimize
the sub-systems, may apply. Rather, within the business world, the
operating conditions of organizations can vary so quickly and so
radically that sometimes such a theorem can be applicable, and some
times it cannot. In the real world, we try to find out in what
circumstances it may or may not apply, and prove it to the client
(sometimes it may be obvious, sometimes it can be difficult). Given the
system and its operating conditions, it would be useful if we could
quickly map this into (1) and (2), and from here derive or deduce the
validity of the theorem you are looking for.

Hope this helps.

Regards,

Alexandre

-------------------------------------
Alexandre J G P Rodrigues
Pugh-Roberts Associates
PA Consulting Group
41 William Linskey Way
Cambridge, MA 02142
USA
Tel: +1 (617) 864 8880
Fax: +1 (617) 864 8884
email: Alex.Rodrigues@PA-Consulting.com
http://www.pa-consulting.com/pra.html

Culture: that knowledge remaining after you have forgotten.

["WALKER, BOB J"]

Jerry poses a very interesting and worthwhile question.

Im not a theoretician but Ive lived the underlying issue for many
years in a commercial context. My practical experience has been quite
clear on the assumption that "If you optimize the performance of the
parts, the whole will also be optimized" The experience is that exactly
the reverse is true, optimizing the parts actually guarantees that the
whole will be sub-optimized, sometimes seriously.

We are used to counter-intuitive results in the SD world and this is a
big one, part of widely held folk wisdom and very pervasive in
management policy setting and decision making. A theoretical
contribution, even a clarification, would be most helpful in supporting
those of us trying to change "the system". Real progress seems to
require more support than the anecdotal evidence most practitioners rely
on.

To set the context... how do we define the "system" of which there can
then be "subsystems". Definitely not trivial unless youre talking about
systems so simple they can genuinely be separated from all other
systems. In my experience these happen only in classrooms and
immediately become irrelevant to the "real" systems we all have to deal
with. The following is drawn from my direct response to Jerry Winston...

In my case the system I want to optimize is Bell Canada, Canadas
largest phone company and the operating component of Canadas largest
Corporate entity. I can even precisely state the optimisation criteria
and measurement... the creation of Shareholder Value computed according
to a rigid set of rules. So far so good.

When I look down and inward, I see that were made up of entities
I could define as systems (say... internal business units). For years
weve wrestled with pushing accountability for performance downward...
usually to the point of carefully designing internal profit centres.
Each is charged with optimizing its performance and "contribution to the
whole".
To our great chagrin I can also attest that this does not and CANNOT
work. Because were so connected and interdependent any optimisation of
a component part
automatically suboptimizes the whole. In SD terms, feedback kicks in and
ruins your carefully designed plans.

This would be true even if there was no cost to the optimisation of the
parts because each would have to accomodate the conflicting interests of
all the others. But there is a cost and it can be huge. For some time I
worked on designing elaborate transfer pricing schemes between internal
subsystems, just to count the beans, in the pursuit of optimisation.
This intervention alone suboptimises the target system because it
consumes energy and money without adding value. In the end, we, like
almost all other telecom companies had to give up. The currently
favoured slogan is "Customer First, One Company, One Team". Sounds
trite, makes good ad copy but really is an explicit recognition of
failure in optimising component systems to the benefit of the whole.

Further, when I look upward and outward from my "defined" system I see
that were part of many other systems. Two in particular would be the
previously mentioned corporate entity and the Canadian
telecommunications marketplace. If the above arguments are true then
optimizing my performance will also act to sub-optimize these
super-systems... but this would be another story....

What I keep coming back to time and again are some of the fundamental
observations of system dynamics... First, complex systems tend to LOW
performance. Second, boundaries between systems are necessary but ONLY
because of our limited, inadequate ability to deal with the real system
of the whole. Any definition of a subsystem is indeed arbitrary and we
deny this only at our peril.

Bob Walker
Director-Performance
Bell Canada
From: "WALKER, BOB J" <bob.walker@bell.ca>

[Joe Kilbride]

> Jerry Winston said:
>
>Around the world, consultants are advising governments that it is
>important to optimise the "performance" of every sub system. This is
>often given as one of the main reasons for "performance measurement":
>to allow the performance of every sub system to be optimised.

Others have responded, citing the dangers of optimising subsystems.
Perhaps we should clarify terms. I can think of two different meanings I
have found in the way people use the term "optimise" in the above
context re: subsystems.

Websters defines "optimise" as "to make as perfect, effective, or
functional as possible." I can interpret that in a couple different ways
and I find that people often do. i.e.,...

1) to some, optimise might mean to MAXIMIZE the subsystem, bring it to
its highest level of performance, independent of the rest of the system

2) to others, optimise might mean to make the subsystem perform in a way
that MAXIMIZES the total system of which it is an interdependent part

(I think) we would all agree that to Optimise sub-systems per #1 above
would be foolhardy and would diminish the performance of the system as a
whole. Goldratt has demonstrated this, as someone else on the listserv
noted.

I have no problem with "optimising", as long as we are clear on what it
is that we are optimising and make certain it is the larger system, not
an interdependent sub-system.

To avoid this confusion around the term "optimise", I often use the
words MAXIMIZE when referring to #1 above, and OPTIMISE when referring
to #2. Another way is to distinguish between local optima (#1 above) and
global optima (#2 above). This is how Goldratt makes the distinction.

In any case, being clear on the THING being optimised is more important
to me and I have experienced confusion around these terms in the past.

Joe Kilbride==jk@mcs.com==Kilbride Consulting,Inc.==630.515.9882
Or visit the KCI web page at http://www.mcs.net/~jk/

[Jim Hines]

I liked Joe Kilbrides clarification of use of term "optimize".

However, I think that we should de-emphasize terms like "optimize" in
favor of the terms like "improve". Our best efforts will not result in
optimized social systems. We can, however, significanlty improve them.

Use of the term "improve" provides the proper standard of comparison for
our work (the current system, not some hypothetical perfect system) and
clues people into when it is O.K. to implement insights: You should
implement policies when you are reasonably sure they will lead to
improvement, you dont need to wait until you can define the perfect
policy.

Jim Hines
LeapTec and M.I.T.
JimHines@Interserv.Com

[Stephen Wehrenberg]

George Richardson wrote:

> I recall hearing one of the authors of "Mankind at the Turning Point" by
> Mesarovic and Pestel (a book published by the Club of Rome after "Limits
> to Growth") claim the same theorem mentioned in this query. Mr. Winston
> may find a reference in that book to the theorem and its source(s).

I have been hoping one of our group woud pursue this. Has anyone tried to
track down the theory via this route? I have a copy of the book, but Im in
the middle of a major office remodeling (self inflicted stress added to the
holiday season!) and the book is packed away in one of 50 or so boxes stacked
in my sunroom.

If I ever get this little construction project under control, Ill take a
look--unless someone already has ...

Steve

--
Stephen B. Wehrenberg, Ph.D.
Chief, Forecasts and Systems, US Coast Guard
Administrative Sciences, The George Washington University
wstephen@erols.com, (202)257-0624
"Born empowered ... what happened?!?"