NetNews Usenet Archive 1992 #30/NN_1992_30.iso/spool/comp/ai/4657

Text File | 1992-12-17 | 7.5 KB | 149 lines

Newsgroups: comp.ai
Path: sparky!uunet!zaphod.mps.ohio-state.edu!rpi!ghost.dsi.unimi.it!univ-lyon1.fr!chx400!news.unige.ch!divsun.unige.ch!afzal
From: [e-mail] (Afzal Ballim)
Subject: Offprints of dissertation on belief systems available
Message-ID: <[e-mail]>
Keywords: Belief, dissertation
Sender: [e-mail]
Reply-To: [e-mail]
Organization: University of Geneva, Switzerland
Date: Thu, 17 Dec 1992 11:34:17 GMT
Lines: 136

Offprints of my Ph.D. dissertation are available to whoever is
interested, simply send my your address. I've included an extended
abstract (1000 words) below:

====================================================================


        ViewFinder: A Framework for Representing,
           Ascribing and Maintaining Nested Beliefs of
                Interacting Agents

                   Afzal Ballim


                 Abstract


Interacting with agents in an intelligent manner means that the computer
program is able to adapt itself to the specific requirements of agents.  The
dissertation is concerned with an important feature necessary for this
ability to adapt:  the use of models of the beliefs and knowledge of the
interacting agents.

The objective of this dissertation is to detail a theory of belief, by which
is meant a theory of how the contents of nested belief models are formed.
The work is motivated by (i) the aspects of representation, formation, and
revision of nested belief models that have been neglected, and (ii) the lack
of a unifying framework for all of these features of nested beliefs.

In much research involving models of the beliefs of agents, the models used
are pre-given.  While this is sufficient in highly constrained domains it is
inappropriate in general.  In more complex domains it is necessary to
dynamically generate these models.  This dissertation is directly concerned
with the problems of dynamically creating such nested models of the beliefs
of agents.


Major Contributions
===================

Using Stereotypes for Belief Ascription:  The use of stereotypes (in
particular, hierarchically organised stereotypes) is shown to be an
excellent method for making shallow belief models, i.e., the system's view
of another agent, but to suffer from serious problems as a mechanism for
making deeper ascriptions.  A precedence inheritance reasoner for
stereotypes is developed as this is necessary in shallow use of stereotypes,
in deep use, and in the fused ascription process later developed.

Perturbation as a Basis for Belief Ascription:  The perturbation model of
belief ascription is investigated, and shown to be excellent for generation
of deep belief models.  However, it relies upon beliefs generally being
common between the ascribing agent (the ascriber) and the agent to whom the
beliefs are being ascribed (the ascribee), and so is inadequate when this
assumption does not hold.

Ascription Operators:  The perturbation method of belief ascription relies
on counter-evidence to block ascription.  In the investigation of
counter-evidence a number of families of ascription operators are
identified.  In particular, a distinction is made between radical operators,
which can form a disjunctive result, and conservative operators which form a
unique result.  Different types of each of these operators are developed.

Weighing Evidence Sources:  The basic perturbation mechanism for belief
ascription intuitively implies a simple recursive algorithm.  This is shown
to be false.  It is demonstrated that pre-stored nested beliefs of the
agents involved in a nesting (and pre-stored beliefs of their views of each
other) must be considered as sources of evidence and counter-evidence for
ascription.  A method for totally ordering these evidence sources is
developed.

A Fused Approach:  The stereotype ascription and perturbation ascription are
fused together via the notion of atypical beliefs .  These are beliefs that
are only held by particular groups of agents.  A mechanism based on lambda
expressions is investigated for fusing atypicality into the perturbation
mechanism.  Evaluation of these lambda expressions in a nested environments
investigated, as to is the nature of transformations on them that is
required when they are ascribed from agent to agent.  Improvements are made
over previously descriptions of these processes.

Interpreted and Ascribed Beliefs:  Interpreted beliefs are those beliefs
attributed to an agent based on inferences from the utterances (or actions)
of the agent.  Ascribed beliefs are those beliefs attributed to an agent
based on principles of commonality and minimal evidence about the probable
background of the agent.  It is claimed that belief interpretation requires
belief ascription as a pre-requisite.  Further, although it might be
expected that interpreted beliefs would always be preferred to ascribed
ones, it is claimed that this is not necessarily the case and that this
largely depends on the inference rules used to interpret the beliefs.  It
follows that interpretation rules must be classed to give some indication of
the reliability of their inferences.

Maintenance of Nested Beliefs:  A number of aspects of belief revision with
respect to nested beliefs are considered.  It is claimed that it is
necessary to distinguish between simulating an agents own revision
mechanism, and making revisions of the belief model of the agent because the
model has proved to be wrong.  In addition, it is shown how either of these
changes can cause the system to want to change its own beliefs, but in
different ways.  Simulating the agents own revision process is more likely
to cause the system to make revisions of its beliefs about the domain, or
topic of conversation, while revisions on the belief model are more likely
to cause the system to make revisions of its beliefs about the agent.  An
operation known as percolation is devised to aid in the former revision
process.

A Framework for Environments:  A general framework for environments is
developed.  The framework is deliberately designed to be open-ended, i.e.,
not completely defined.  A number of important types of environment are
discussed.  It is shown how environments may be used as a mechanism for
reasoning about ordering relations, and how, thus, they provide an
appropriate medium for reasoning about different ordering relations, such as
confidence relations, inheritance leaning relations, etc.  Particular
attention is paid to the representational problems discussed in the
background section.

Ascription Operators as Environment Projection:  The ascription operators
previously developed are generalised to be operators that cause projection
of one environment on another.  Environment projection operators are studied
in more detail, and a number of extra projection operators are proposed,
including Bayesian projection, Fuzzy projection, and Logic Foundational
projection.

Environment Projection as a Fundamental Operator:  It is shown that
environment projection can be seen as a fundamental operator underlying many
important processes in AI, including belief ascription, inheritance
reasoning, truth maintenance, belief revision, merging of intensional
descriptions, and metaphor generation.  An environment framework may thus
serve as a basis for investigation, development, and implementation of all
of these processes.



--
Afzal Ballim                 |BITNET,EARN,MHS,X.400: [e-mail]
 ISSCO, University of Geneva |UUCP: mcsun!divsun.unige.ch!afzal
 54 route des Acacias         |JANET: [e-mail]-relay
 CH-1227 GENEVA,Switzerland  |CSNET,ARPA: [e-mail]