Copy Link
Add to Bookmark
Report
Machine Learning List Vol. 2 No. 14
Machine Learning List: Vol. 2 No. 14
Monday, August 13, 1990
Contents:
Chess vs. Structural Design [ML List: Vol. 2 No. 13]
Special Issue of MLJ on Symbolic Learning and Robotics
Symposium on Learning Methods for Planning and Scheduling
The Machine Learning List is moderated. Contributions should be relevant to
the scientific study of machine learning. Mail contributions to ml@ics.uci.edu.
Mail requests to be added or deleted to ml-request@ics.uci.edu. Back issues
may be FTP'd from ics.uci.edu in /usr2/spool/ftp/pub/ml-list/V<X>/<N> or N.Z
where X and N are the volume and number of the issue; ID & password: anonymous
------------------------------
From: Lou Steinberg <lou@atanasoff.rutgers.edu>
Subject: Re: Chess vs. Structural Design [ML List: Vol. 2 No. 13]
Date: 18 Jul 90 15:55:08 GMT
I have to both agree and disagree with Yoram Reich's claim that
structural design is unlike chess as a learning domain.
I agree with everything he says characterizing structural design as a
problem with uncertainty even at the very base level - e.g., in real
world design problems, at EVERY level of description, the goal (i.e.
the specifications) and the effects of the operators (e.g. the
behavior of a given structure) are only known approximately.
On the other hand, this isn't the whole story. It is sometimes both
possible and useful while doing a SUBTASK of design to PRETEND that
you do indeed have complete knowledge of the goal, operator, etc.
Take digital VLSI circuit design as an example - in the real world
goals include things like speed and silicon area, and these goals are
known at best approximately. Furthermore, digital circuits don't
behave in a perfectly digital manner - "digital" is just an
approximation. However, consider a subtask like "find the shortest
set of paths you can for wires from these points to those points".
For this problem, we can pretend we know exactly what the goals and
operators are, and write a program which solves it. In fact, we are
lying when we pose the problem this way, because in fact it is
physically possible not only to route a wire but to also move the
things the wire is connecting, and doing so may lead to a better
circuit. Furthermore, the manufacturing process won't put the wires
exactly where we say to put them. However, it is still useful to
write this program, because it lets us get an acceptable design done
with a feasible amount of effort, and other parts of the design and
manufacturing process handle the fact that what this one program does
is based a little bit on a lie.
Cf. also the work by Jack Mostow which views knowledge compilation
from specifications as the inner loop of a larger design process
which includes changes in the specifications, e.g.
Jack Mostow, "Towards automated development of specialized
algorithms for design synthesis: knowledge compilation as an
approach to computer-aided design", Research in Engineering
Design, V. 1 No. 3, 1989. Also available as Rutgers AI/Design
Project Working Paper Number 141
Thus, while design as a whole is unlike chess in this way, there are
useful subtasks of design that are like chess in this way.
(By the way, I am afraid there may be some misunderstanding of the
term "structural" design. As I use the term (actually I tend to say
"structure" design), and as I think Tom Dietterich was using it, the
term doesn't mean "design of civil engineering structures like
buildings and bridges", it means "design of an artifact where we don't
know when we start a given problem what pieces the artifact will have
or how they will connect/interact". This is to distinguish it from
"parameter" design where we do know the pieces and interactions, but
need to determine parameters like sizes and materials.)
Oh, and one further quibble. Yoram says: "Even good optimization
techniques do not scale up to real problems." In fact, a system
combining traditional numerical optimization, genetic learning
algorithms, and expert-system-style rules is actually in production
use for parameter design of real jet aircraft engines and a whole
bunch of other things at GE. The system doing this is Engineous,
developed by Siu Tong and others at GE.
--
Lou Steinberg
uucp: {pretty much any major site}!rutgers!aramis.rutgers.edu!lou
arpa: lou@cs.rutgers.edu
------------------------------
Date: Thu, 9 Aug 90 12:30:17 PDT
From: Tom Dietterich <tgd@turing.cs.orst.edu>
Subject: Structural Design
I am responding to Yoram Reich's refutation of my claim that
structural design provides a good domain for knowledge compilation
research. Reich's main objections are that design problem solving is
complex and that particularly the problem formulation (i.e., the
initial specification) is generally imprecise, self-contradictory,
and subject to modification during the design process.
Of course I agree with these points in general. Engineering design is
very complex, and I've been studying it for only the past 5 years.
However, there are small subproblems in mechanical design where it is
possible to provide precise specifications and objective functions so
that the design process does become completely well-specified. In
such cases, the only problem is computational, and it is appropriate
to apply knowledge compilation methods.
My student Giuseppe Cerbone and I are developing knowledge compilation
methods for the following small subproblem:
Given: A set of static loads (locations, magnitudes, and directions)
A set of stable connection points
A collection of convex forbidden regions
Densities of tension and compression members.
Find: A minimum weight statically determinate structure containing
only pure tension and pure compression members that stabilizes
all loads and does not enter any forbidden region.
The problem can be cast as a constrained non-linear optimization
problem, and locally optimal solutions can be found by numerical
methods. Our goal is to find fast, nearly-correct methods that scale
to large problems.
Now a general theory of mechanical design will need to address many
problems that arise from imprecise specifications and incomplete
knowledge, but I still believe that there will be many subproblems
like this one in which purely computational solutions can be found.
These computational solutions can be embedded in a larger man/machine
framework: people can propose specifications and the system can
optimize for those specs. Then, if the specs change, the system can
re-optimize for them. Indeed, one of the potential payoffs of
knowledge compilation research is that it will provide much more
insight into the solution space than current numerical optimization
methods provide.
Reich says "There is NO theory of systematic synthesis in structural
design." This is true, and it is our goal, at least for this small
subproblem, to develop such a theory.
------------------------------
Date: Fri, 27 Jul 90 08:46:11 EDT
From: John Laird <laird@caen.engin.umich.edu>
Subject: Special Issue of MLJ on Symbolic Learning and Robotics
Call for Papers
Special issue of Machine Learning Journal on Symbolic Learning and Robotics
Guest Editor: John Laird
Deadline: November 15, 1990
More information: laird@caen.engin.umich.edu
------------------------------
From: Steven Minton <MINTON@PLUTO.ARC.NASA.GOV>
Subject: Symposium on Learning Methods for Planning and Scheduling
Call for Participation
Symposium on Learning Methods for Planning and Scheduling
Palo Alto, California
January 5-6, 1991
Research on planning and scheduling systems has shown that domain
knowledge is crucial for effectively coping with complex, changing
environments. Recent advances in machine learning have provided a
deeper understanding of learning mechanisms relevant to acquiring such
knowledge. However, most current systems are considered successful
from either the planning and learning perspectives, but not both. The
development of practical systems that can modify and improve their own
behavior when confronted with new or changing environments remains a
significant challenge.
On January 5 and 6, 1990, a symposium will take place in Palo Alto,
California, to address the application of learning techniques to
planning and scheduling problems. The meeting will bring together
active researchers on this important topic, letting them report their
recent results and discuss unresolved problems. It will also establish
communication among researchers from different paradigms, enabling
systems to be developed that respond to real planning and scheduling
problems.
The ideal participant will be an active researcher who has
has worked in the intersection of learning and planning and/or
scheduling, including:
- Learning to improve the efficiency of planning
- Acquisition of planning and execution skills
- Organization and indexing of plan knowledge
- Learning to plan in changing environments
- Computational models of human plan acquisition
All speakers and participants will be invited. Applicants
interested in either speaking or participating should send a short
research summary (maximum of three pages) describing their research
efforts and interests by October 1, 1990, to the program chair:
Steven Minton
AI Research Branch, MS 244-17
NASA Ames Research Center
Moffett Field, CA 94035,
USA
Electronic mail: Minton@pluto.arc.nasa.gov
Telephone: (415) 604-6522
Travel support for participants may be available; we will have more
information on support by the submission deadline. For further
information contact the program chair at the above address.
This symposium is the third in an annual series of meetings
organized by the Institute for the Study of Learning and Expertise, a
nonprofit organization devoted to encouraging research and education
in machine learning and related disciplines. The results of the 1989
and 1990 symposia, which focused on scientific discovery and concept
formation, will soon appear in collected volumes published by Morgan
Kaufmann.
------------------------------
END of ML-LIST 2.14
