The soft versus hard debate

Paul J. Werbos (pwerbos@nsf.gov)
Mon, 26 Oct 1998 05:32:26 +0100 (MET)

It has been very interesting to watch this AI versus fuzzy logic (or "hard"
versus "soft" debate).

On the one hand, I am extremely impressed by the quality and results of
Lotfi's effort to defend fuzzy logic.
The neural network community faces many of the same issues that the fuzzy
logic community does, but we have clearly suffered a great deal from not
having anyone to move in and play the kind of role that Lotfi has played in
defense of the fuzzy world.

There are many reasons for this. Ironically, one reason is that some of us
oldtimers in the neural world -- Bernie Widrow and myself -- tend to shy
away from extreme binary debates. (There are also a few folks who thrive on
binary debates, but focus mainly on defending their own substream of the
neural net field.) When there are well-worked out opposing positions,
engaged in stark black-and-white debates... we have a tendency to look for
the truth somewhere in the middle, rather than the extremes. We expect the
truth to be in the middle -- but we also expect it to be messy and complex,
to some degree, like a vivid kaleidoscope picture rather than a fuzzy
gradient. However, when we try to jump directly to that truth in the
middle, we are not always meeting the needs of the community as a whole.
Someone needs to keep the basic issues and motivations alive, to permit the
energy to flow in the end to the correct position in the middle. The issues
here involve not just truth, but human motivations and excitement. Again, I
must say that I envy the fuzzy world for having someone like Lotfi to fill
in this crucial need.

On the other hand, this is only part of what we need. We also do need
people who are willing to go past the stage of debate, and do the messy
hard mathematical work needed to build a more solid bridge between the
islands of understanding reflected in different intellectual disciplines.

One of these areas where new work is critically needed -- to make the
middle position SOLID --
concerns the relation between soft computing and control engineering. In
fact, some people point out that the delivery of working products in
control is perhaps hundreds of times as large or larger than the delivery
in AI. Why chase after a wounded mouse? On the other hand, AI people would
say we should pay attention to them because they are curators of an
important intellectual issue: how can we build systems which are TRULY
intelligent,
in the same way that brains are intelligent?

To really address the issues of this debate in an honest way... part of our
job is to do more work, ourselves, to address the two legitimate questions
raised by the control engineering and AI communities:

(1) How can we arrive at the kinds of stability guarantees necessary for
the deployment of our new designs in
`real-world control applications, particularly given the stringent safety
requirements of many applications?

(2) How can we address the higher-order (learning-based) planning and
decision-making capabilities which are the hallmark of real intelligent
systems like brains? (The AI people vacillate a lot on teh issue of
learning, but they certainly would say -- legitimately -- that it's not
"intelligent" if the behavior is all based on preprogrammed IF-THEN rules.
In any case, the onus is on us to address the real problem, where LEARNING
ability is the very center.)

Of course, Sugeno has done a magnificent service to the fuzzy community in
addressing the first of these two questions, and helping inspire a much
larger community of people doing the same thing, generating practical
applications. Without him, I doubt the fuzzy world would have gotten where
it is today. He has received a great deal of respect -- but perhaps he
deserves even more.

BUt there is more work to be done. For true intelligent systems -- to
address the second question, which Minsky and Simon have often posed very
effectively (often, but not always) -- we must develop learning or adaptive
decision-making and control systems.

There has been great progress in the fuzzy world, recently, in adding
adaptation. The fuzzy-neural communication,
encouraged by Lotfi's talk about "soft computing," has been one factor in
helping this process. (I like to believe that some of the earlier
suggestions by Ron Yager and myself may have helped as well.) Yet, in the
world of decision-making, the vast bulkl of this work has all involved a
species of control which I would call
"cloning" or "supervised control." People like Minsky look at this stuff,
and they laugh, for good reason. It can be very useful in practice, but it
is not at all true intelligence, for the following reason. In supervised
cloning, the learning is all based on "a teacher." Someone TELLS the fuzzy
or neural system what its output SHOULD BE, and trains it to learn the
static input-output mapping (in most cases) or the dynamic input-output
mapping (in a few cases).
BUt the adaptive system does not learn how to find the best output by
itself; it can only copy someone who already knows. It cannot be creative,
and devise a new strategy of action; it can only copy someone else's
strategy.

IN my limited knowledge, Hamid Berenji is the only person in the fuzzy
world who has gone beyond this static framework, and beyond the notions of
simple setpoint control, to devise fuzzy learning systems which really
address the problem of planning over time, which is critical to real
intelligent systems. I am sujrprised that this important work has not been
followed up on to a greater degree. It is part of the correct answer to the
complaints of the AI people... but it is our fault we have not developed
that answer further, to make it more complete and more available.

In fact, Berenji's workl is part of a larger trend in the soft computing
world, which the AI world has PARTIALLY accepted and use... a trend which
is sometijmes called Approximate Dynamic Programming (ADP), sometimes
called adaptive critics (a term invented by Bernie Widrow, the first person
to implement such a system -- though most of the AI people studiously
neglect him and other aliens to their world), and sometimes called
reinforcement learning.

BUt: the classical control people have asked agian and agian: where are
your stability guarantees for such learning systems? "We have lots of
theorems for classical adaptive control; where are your guarantees here?"

In fact, it now appears -- as of this year -- that we may be able to get
much STRONGER and more REALISTIC guarantees of stability for the
reinforcement learning designs than are available with the traditional
adaptive control. This requires a strategy of cooperation with that
paradigm, rather than debate. It also requires lots of work. I have done
some of the preliminary work myself. (For details, go to xxx.lanl.gov, in
the nlin-sys area,
and look at adap-org/9810001.) BUt theer is a lot more to be done. I hope
that some of you might consider
participating in that task.

For the record, these reinforcement learning designs are general
mathematics. They do not come with a "classical" or "neural" or "fuzzy"
color. They can be applied to adapt any of those structures, if they are
differentiable and
well-parametrized. (e.g. One must use elastic fuzzy logic instead of the
usual piecewise-linear stuff where only
membership functions are adapted.) So far as I know, however, no one in the
fuzzy world has ever tried to do those kinds of learning tasks (with the
partial exception, again, of Berenji).

Sorry to be so long... but these are simply not trivial issues; it is
essential to move beyond binary debates to
building solid, detailed worked-out bridges between these various
communities. I hope you can help.

Best of luck,

Paul W.

P.S. These are NOT official NSF views!!! However, the general idea of
funding work which builds serious technical bridges between paradigms, and
advances information technology in a general way -- especially in support
of learning -- is certainly compatible with the new kinds of priorities we
have been looking at here. The program I co-manage is now renamed "Control,
networks and computational intelligence." We hope to update our web pages
(under www.nsf.gov) to reflect these broad new priorities in the near future.

.

############################################################################
This message was posted through the fuzzy mailing list.
(1) To subscribe to this mailing list, send a message body of
"SUB FUZZY-MAIL myFirstName mySurname" to listproc@dbai.tuwien.ac.at
(2) To unsubscribe from this mailing list, send a message body of
"UNSUB FUZZY-MAIL" or "UNSUB FUZZY-MAIL yoursubscription@email.address.com"
to listproc@dbai.tuwien.ac.at
(3) To reach the human who maintains the list, send mail to
fuzzy-owner@dbai.tuwien.ac.at
(4) WWW access and other information on Fuzzy Sets and Logic see
http://www.dbai.tuwien.ac.at/ftp/mlowner/fuzzy-mail.info
(5) WWW archive: http://www.dbai.tuwien.ac.at/marchives/fuzzy-mail/index.html