Copy Link
Add to Bookmark
Report
AIList Digest Volume 4 Issue 069
AIList Digest Tuesday, 8 Apr 1986 Volume 4 : Issue 69
Today's Topics:
Seminars - Metaplanning: Controlling Planning in a Complex Domain (CMU) &
Rule-Based Systems and Heuristic Classification (SU) &
The MACE System (USC) &
Expert Systems for System Management (MIT) &
Temporal Theorem Proving (SRI) &
Network Propagation for Reasoning about Uncertainty (CMU) &
Optical Artifical Intelligence Research in ECE (CMU) &
Pragmatic Modeling: Robust NL Interface (MIT)
----------------------------------------------------------------------
Date: 1 April 1986 0108-EST
From: Paul Birkel@A.CS.CMU.EDU
Subject: Seminar - Metaplanning: Controlling Planning in a Complex Domain (CMU)
Metaplanning:
Controlling Planning in a Complex Domain
Dissertation Proposal
Friday, April 4th
1:00-2:30 PM
Wean Hall 5409
All planners metaplan; few do so explicitly. Many planners find very
simple control mechanisms sufficient; the added overhead of a metaplanner
outweighs any apparent advantages. Whether implementing an explicit
metaplanner increases the capabilities of the resulting system is unknown.
Complex domains, such as therapy planning, include problems which would
be best handled by a metaplanner identifying and choosing alternative
planning strategies separate from the process of plan generation. These
problems include: unresolvably conflicting goals, conflicting measures of
goal satisfaction, unreliable operators, and incompletely specified initial
states. Previous therapeutic (@b<MYCIN>, @b<ONCOCIN>) and non-therapeutic
(@b<NOAH>, @b<SIPE>) planners alike are incapable of explicitly reasoning
about, and solving, combinations of these types of problems. A hierarchical
therapeutic planner will be implemented based on a @b(MOLGEN/SPEX) hybrid
architecture incorporating both tactical planning and strategic metaplanning
components. Four additional planning techniques are proposed which will be
developed and integrated into the architecture. The metaplanner will
subsequently be extended to achieve acceptable clinical performance on two
dozen clinical cases covering all combinations of these problems. The
performance of the system with and without the planning extensions, and
with and without the metaplanner will be analyzed.
A copy of the thesis proposal is available in the
CS lounge, 4th floor, Wean Hall. Please contact me
for additional copies of the proposal (its long!).
birkel@a or x3074
------------------------------
Date: Mon 31 Mar 86 18:31:41-PST
From: Christine Pasley <pasley@SRI-KL>
Subject: Seminar - Rule-Based Systems and Heuristic Classification (SU)
CS529 - AI In Design & Manufacturing
Instructor: Dr. J. M. Tenenbaum
Title: Rule-based Systems; Application to Heuristic Classification
Speaker: William J. Clancey
From: Knowledge Systems Laboratory
Date: Wednesday, April 2, 1986
Time: 4:00 - 5:30
Place: Terman 556
This talk provides an broad overview of expert systems research,
using the Neomycin program as an example. We consider in particular the
rule-based knowledge representation, showing how rules can be controlled
by an inference procedure. Generalizing from this example, we consider
first the heuristic classification method of problem solving, showing how
a broad range of well-structured problems--embracing forms of diagnosis,
catalog selection, and skeletal planning--are solved in typical expert
systems. Next, we consider kinds of problems that expert systems can be
used to solve, emphasizing the idea of a "system in the world" that is being
synthesized or analyzed. Finally, we introduce the idea of a qualitative
model, showing how different kinds of network formalisms are used in expert
systems to describe processes. The material in this talk will enable you
to relate the kinds of problems, solution methods, and representations in
expert systems.
------------------------------
Date: 2 Apr 1986 18:45-EST
From: gasser@usc-cse.usc.edu
Subject: Seminar - The MACE System (USC)
USC DPS GROUP MEETING
Wednesday, 4/9/86 3:00 - 5:00 PM
Seaver Science 319
Les Gasser will speak on the MACE system.
MACE is a testbed for building generic Distribiuted AI systems from
organized collections of active "intelligent" entities called
@i[Agents] which run in parallel. It comprises a language for
describing agents, a language for describing a network of processors
upon which the agents run, and a simulator for executing the agents
in parallel. This talk will describe the philosophy and design goals of
MACE, the current versions of the MACE description languages, the
MACE simulator, and briefly discusses several experimental MACE
implementations.
The MACE language is constructed in two parts: the MACE Agent Description
Language which is sufficient for expressing agents or collections of
agents at any level (including composite agents), and the MACE
Environment Description Language which describes the underlying
computation hardware and simulator parameters. Individual
agents may draw upon other existing languages.
MACE has been implemented in COMMON LISP on a TI Explorer Lisp
Machine. We have several trial systems implemented (*) or partially
implemented (-).
- An ACTORS-like recursive Fibonacci computation which
we have tested by creating up to 90 agents running in parallel.*
- An agent called BUILDER which interactively builds other agents
through a second agent called USER-INTERFACE, both agents running
in parallel.*
- An agent-based production system where each rule is an agent, and
there is no global database nor centralized inference engine. (-)
- An 8-node hypercube with MACE agents running on each node, and a
parallel broadcast facility among agents.*
- A distributed, multi-level blackboard built of agents. (-)
- A two-robot cooperative planner. (-)
Questions: Dr. Les Gasser, (213) 743-7794, gasser@usc-cse.usc.edu
------------------------------
Date: Wed 2 Apr 86 08:54:20-EST
From: Natalie F. Tarbet <NFT@XX.LCS.MIT.EDU>
Subject: Seminar - Expert Systems for System Management (MIT)
[Forwarded from the MIT bboard by SASW@MIT-MC.]
Fourth in a series of seminars on Large and Complex Computer
Systems in the Commercial World
"Expert Systems for System Management and Control
or
What jobs in a large computing center can be automated?"
Keith R. Milliken
IBM, Thomas Watson Research Center
Yorktown Heights, NY
NE43-512A
Wednesday, April 2, 1986 at 3:15 p.m.
Several years ago, IBM's Thomas Watson Research Center began to
develop an expert system to assist with the operation of a large
computing complex. This expert system, called YES/MVS (Yorktown
Expert System / MVS Manager), runs in real-time and can either
give advice or automatically take actions to manage computing
resources and respond to problems in a running system. This system
is of interest because it actively helps control, in real-time,
a very complex process. YES/MVS has been used extensively in the
Yorktown Computing Center, and a second version is now being developed.
We will briefly describe YES/MVS and then focus on some of the expert
system issues that have arisen during YES/MVS development and the
approaches taken to resolve them. Two of the issues that will be
emphasized are (1) knowledge representation for process control
expert systems and (2) approaches to knowledge base organization
that reduce the difficulty involved in modifying a large knowledge base.
The latter issue is especially important in the automation of computing
system operation because there are large variations between computing
centers in operational policy.
We shall briefly describe related efforts to automatically analyze the
performance of large computing systems, to deveop a special purpose
shell for computer performance expert systems and to use rule-based
techniques to control resource allocation in a large computing system.
Host: Arvind
------------------------------
Date: Wed 2 Apr 86 17:21:06-PST
From: LANSKY@SRI-AI.ARPA
Subject: Seminar - Temporal Theorem Proving (SRI)
TEMPORAL THEOREM PROVING
Martin Abadi (MA@SAIL)
Stanford University
11:00 AM, MONDAY, April 7
SRI International, Building E, Room EJ228 (new conference room)
In spite of the wide range of applications of temporal logic,
proof techniques (especially for first-order temporal logic (FTL))
have been quite limited up to now. We have developed a proof system R
for FTL. The system R is based on nonclausal resolution; proofs are
natural and generally short. Special quantifier rules, unification
techniques, and a resolution rule are introduced. The system R is
directly useful for such tasks as verification of concurrent programs
and reasoning about hardware devices. Other uses of temporal resolution,
such as temporal-logic programming, are currently being considered.
We relate R to other proof systems for FTL and discuss completeness issues.
In particular, one variant of R is ``as complete as'' an extension of Peano
Arithmetic. We also describe resolution systems analogous to R for other modal
logics. In fact, the resolution techniques and the corresponding completeness
arguments apply to a large class of modal logics.
------------------------------
Date: 2 April 1986 1720-EST
From: Betsy Herk@A.CS.CMU.EDU
Subject: Seminar - Network Propagation for Reasoning about Uncertainty (CMU)
Speaker: Judea Pearl, UCLA
Date: Tuesday, April 15
Time: 3:30 - 5:00
Place: 5409 Wean Hall
Title: Network propagation for reasoning about uncertainty
Abstract:
In order to meet requirements of modularity, transparency and
flexibility, the designers of 1st-generation expert systems have
abandoned traditional probability theory and ventured to devise new
formalisms for managing uncertainties. The talk will describe a
message-passing scheme in propositional networks which, using
traditional probability theory, fulfills these objectives of
expert systems technology.
I will argue that the notion of TRANSPARENCY is closely related to
reasoning with GRAPHS, namely, that an argument is perceived to be
"psychologically meaningful" if its derivational steps correspond
to mental tracings of pre-established links in some conceptual
dependency network. Accordingly the first part of the talk will
introduce an axiomatic legitimization of representing inferential
dependencies by networks, and will compare the properties of two
such representations: Markov Networks and Bayes Networks.
The second part will introduce a calculus for performing inferences
in Bayes Networks. The impace of each new evidence is viewed as a
perturbation that propagates through the network via asynchronous
local communication among neighboring concepts. We show that such
propagation mechanism facilitates flexible control strategies and
sound explanations, that it supports both predictive and diagnostic
inferences, that it is guaranteed (in sparse graphs) to converge in
time proportional to the network's diameter, and that every
proposition is eventually accorded a measure of belief consistent
with the axioms of probability theory.
------------------------------
Date: 3 April 1986 1023-EST
From: Richard Wallstein@A.CS.CMU.EDU
Subject: Seminar - Optical Artifical Intelligence Research in ECE (CMU)
Robotics Seminar
3:30 Friday April 11, 4623 Wean Hall
David Casasent, Director
Center for Excellence in Optical Data Processing
Department of Electrical and Computer Engineering
OPTICAL ARTIFICAL INTELLIGENCE RESEARCH IN ECE
Optical feature extraction and correlation distortion-invariant multi-class
multi-object recognition and identification research will be reviewed. This
will be followed by a discussion of optical artificial intelligence efforts
currently in progress. This effort includes: optical relational graph and
decision net processors, optical symbolic processors, optical associative
memory processors, and optical neural net processors.
------------------------------
Date: 4 Apr 1986 09:57-EST
From: Brad Goodman <BGOODMAN at BBNG>
Subject: Seminar - Pragmatic Modeling: Robust NL Interface (MIT)
[Forwarded from the MIT bboard by SASW@MIT-MC.]
BBN Laboratories Inc.
Science Development Program
AI/Education Seminar
Speaker: Professor Sandra Carberry
University of Delaware
Title: Pragmatic Modeling: Toward a Robust Natural
Language Interface
Date: Tuesday, April 15th, 10:30 a.m.
Place: 2nd floor large conference room
BBN Labs, 10 Moulton Street, Cambridge
PRAGMATIC MODELING:
TOWARD A ROBUST NATURAL LANGUAGE INTERFACE
Naturally occurring dialogue is both imperfect and incomplete. Not
only does the information-seeker fail to communicate all aspects of his
underlying task and partially constructed plan for accomplishing it, but
also his utterances are often imperfectly or imcompletely formulated. It
appears that human information-seekers expect an information-provider
to facilitate a productive exchange by assimilating the dialogue and
using this knowledge to remedy many of the information-seeker's faulty
utterances.
This talk will describe an on-going research effort aimed both at
developing techniques for inferring and constructing a user model from
an information-seeking dialogue and at identifying strategies for using
this model to develop more robust natural language interfaces. Emphasis
will be on the dynamic construction of the task-related plan
motivating the information-seeker's queries, and its application
in handling pragmatically ill-formed and incomplete utterances.
------------------------------
End of AIList Digest
********************
