🚀 go-pugleaf

                 C A L L  F O R  P A R T I C I P A T I O N S

 AI'94 Tutorial on Knowledge Acquisition and Maintenance with Ripple Down Rules

				 by

                              P.Compton 
               School of Computer Science and Engineering, 
               University of New South Wales, Sydney 2052,
                      email compton@cse.unsw.edu.au 


GOALS

The aim of this tutorial is to teach participants a simple but powerful
knowledge acquisition (KA) technique for building knowledge based systems
(KBS), Ripple Down Rules (RDR).  

The problems in KA arise partly because of our naive cultural expectations
as to what an expert is capable of providing.  The tutorial will encourage
participants to a critical view as to what knowledge acquisition is about
and provide a standpoint for understanding and evaluating KA solutions
proposed by current research.  

RDR will be proposed as a partial solution to the KA problem.  It is
intended that the tutorial will leave participants with a sufficiently
detailed understanding of RDR, its strengths and limitations, so that
without further KBS background they will be able apply RDR to building KBS.

It is intended that,  assuming reasonable programming skills, participants
will be able to build their own RDR interpreters, adapt existing expert
system shells to use RDR, or extend existing information systems to include
RDR, if desired.  It will be possible to present the algorithms involved in
sufficient detail for this because of their simplicity.  However, no prior
experience will be assumed and participants will be provided with RDR
software sufficient to build reasonably large systems.

Finally it would be hoped that participants would get a clear idea of the
limitations of RDR and perhaps be encouraged to explore solutions to these
problems.


BACKGROUND

The first major problem in the successful development of a KBS is the
choice of an application.  Assuming the application is suitable, to build a
significant KBS requires considerable effort, particularly in acquiring and
maintaining the expert knowledge needed by the system.   This has been
described as the "knowledge engineering bottleneck".  The essential problem
in KA is that experts never explain how they reach a conclusion, rather
they justify that their conclusion is correct.  This justification depends
on the context in which it is given, whether the person for whom the
justification is intended is a lay person, a trainee and expert etc.  That
is, knowledge is always "situated" in a context.  Further, it is not part
of an expert's normal skill to be able tocategorise and describe the type
of problem solving they carry out.  

Most current research aims at setting up methodologies to approach these
problems in a disciplined and careful way.  Such methods, in common with
conventional software engineering techniques, are powerful, interesting,
have wide applicability, but require sophisticated practitioners.   Ripple
Down Rules is knowledge acquisition methodology which assumes that since
knowledge is given only in a specific context it should be used only in the
same context.  This leads to an approach to building KBS by incremental
refinement covering maintenance as well as initial development.  It also
allows the knowledge being added to be easily validated.  This approach has
a number of consequences.  

Firstly, RDR allow experts to build KBS without any knowledge engineering
or programming assistance or training.  This has been demonstrated in the
PEIRS system at St.Vincent's Hospital Sydney.  This systems provides
interpretative comments for diagnostic pathology laboratory reports.  It
currently has over 2000 rules and is 95% correct in the domains it covers.
It would seem to be one of the largest medical systems in actual routine
use.  It also seems to be the only large system in use built without a
knowledge engineer (or without induction).  Importantly it was put into
routine use with only 200 rules and all other knowledge has been added
while the system has been in use as a minor extension to  the expert's
normal duties.

Secondly, RDR blur the distinction between maintenance and initial
development.  The cost of adding a rule with an RDR system is the same
regardless of how large the system is and how long it has been under
development.  The issue of loss of expertise and the difficulty of dealing
with a large structure built by someone else does not arise in the
incremental approach used with RDR.  

Thirdly RDR seem to be the first practical approach to KBS based on a
situated cognition perspective.  Situated cognition is increasingly the
basis of a critique of traditional  approaches to AI.  RDR eschew the
notion that it is necessary to search for the best or right way of
constructing a KB.  Rather knowledge is constructed in context as a
consequence of applying insight to a concrete problem.  RDR thus emphasise
test and repair rather than problem solving and domain analyss.

RDR do not immediately solve all problems and have their own intrinsic
problems.  The most important problem is that the KB may contain repeated
knowledge acquired through repeated KA of the same knowledge.   Brian
Gaines (University of Calgary) has developed an inductive method of
building or compressing an RDR KB which will be described in detail. 
However it has also been shown that this is minor rather than major problem
(PEIRS has not yet been compressed).  Secondly, RDR were developed for
applications where a single classification is required and all the data are
available before inference. Extensions to RDR to deal with multiple
classifications and problems such as configuration will be described. 
Participants should expect to gain some insight into how these problems can
be dealt with, but the major focus of the tutorial will be on standard
single classification RDR.  Finally, the development of  domain models and
suitable data abstraction for RDR will be discussed.  


TUTORIAL OUTLINE

KA problems
     KBS case study.  
     KA problems and philosophy. 
     Selecting a KBS application
     Current KA research

Ripple Down Rules. 
     Principles, algorithms, knowledge representation.  
     PEIRS.  
     RDR software demonstration

Problems and Extensions
     RDR and Induction.  
     Multiple Classification RDR.  
     Configuration.  
     Data abstraction


TUTORIAL MATERIALS

Participants will be provided with notes and/or copies of key papers on RDR
and a demonstration program for Apple Macintosh or Windows on a PC.  The
demonstration programs are stand alone and will enable participants to
build significant RDR systems.  Their best use however will be as a
demonstration program as RDR are best used embedded within an information
system.  The software will be demonstrated during the tutorial.  It is not
necessary for participants to bring a portable Macintosh or PC to the
tutorial, but it will facilitate participation.


BIO-DATA OF THE PRESENTER

Paul Compton is an Associate Professor in the School of Computer  Science
and Engineering, University of New South Wales.  Before coming to UNSW in
1990, he was Head of Computing Services at the Garvan Institute of Medical
Research, Sydney.  Together with Kim Horn and Ross Quinlan he was
responsible for GARVAN-ES1 one of the first medical expert systems to reach
routine use.  This system is also one of the very few examples of KBS
maintenance as it was continuously maintained over five years.  RDR grew
out of this maintenance experience. He is a member of the editorial board
of the Knowledge Acquisition Journal, has been on the organising committee
of various international Knowledge Acquisition Workshops and is a co-chair
of the Japanese Knowledge Acquisition Workshop later this year.  


-------------------------------------------------------------------------------
       Seventh Australian Joint Conference on Artificial Intelligence (AI'94) 

                       STRUCTURED SEQUENCE TUTORIALS

A structured sequence of pre-conference tutorials on several aspects of
applied AI has been organised. Tutorial participants will be able to select
from a choice of tutorials to suit their specialist requirements.

The following are the list of tutorials organised for AI'94. All
participants of any of these tutorials may attend the talk entitled
"Introduction to Artificial Intelligence". This is a complementary session
for all tutorial participants.
                    

Guide:
(y) - indicate YES
(n) - indicate NO
(t) - indicate THEORETICAL SESSION
(p) - indicate PRACTICAL SESSION
(d) - indicate DEMONSTRATION SESSION
1hr - indicate one hour
4s  - indicate four sessions
3s  - indicate three sessions 

Note: Each session is One and a Half hours long.


--------------------------------------------------------------------------------
No.     Tutorial Title               Presenters            CODE Length Practical
--------------------------------------------------------------------------------
[0] Introduction To Artificial       Not Confirmed yet     AI     1hr      -
    Intelligence

[1] Constraint-Based Reasoning       Dr. H. W. Guesgen     CBR    4s       y

[2] Fundamentals of Fuzzy Logic      Dr. A. Sekercioglu    FLFC   3s       y
    and Fuzzy Logic Controllers      G.K. Egan
               
[3] An Introduction to Evolutionary  Dr. X. Yao            EC     3s       y
    Computation                      Prof Z. Michalewicz
               
[4] Building Intelligent Decision    Dr. J. Zeleznikow     IDSS   3s       y
    Support Systems through the use  Mr. Dan Hunter
    of multiple reasonig Strategies
                
[5] Intelligent Learning Database    Dr Xindong Wu         ILDB   3s       y 
    Systems            

[6] Nonmonotonic Reasoning           Dr. M-A Williams      NMR    3s       y
                                     Dr. G. Antoniou

[7] Theoretical Foundations of       Dr. A. G. Hoffman     TFML   3s       n
    Machine Learning                 Dr. Shyam Kapur
                                     Dr. Arun Sharma

[8] Knowledge Acquisition and        Dr. P. Compton        KAM    3s       d
    Maintenance with Ripple Down 
    Rules     

[9] Hybrid (AI symbolic,             Dr. Nik K. Kasabov    HS     3s       d
    Connectionist, Fuzzy, Chaotic) 
    Systems


--------------------------------------------------------------------------------
Tentative Tutorial Timetable
--------------------------------------------------------------------------------

Monday 21/11/94 (Day 1)
=======================
 9.30 - 10.30:  Introduction to AI
10.30 - 11.00:  Morning Tea Break
11.00 - 12.30:  NMR(t)  IDSS(t) EC(t)   TFML(t)
12.30 -  2.00:  Lunch
 2.00 -  3.30:  NMR(t)  IDSS(t) EC(t)   TFML(t)
 3.30 -  4.00:  Afternoon Tea Break     
 4.00 -  5.30:  NMR(p)  IDDS(p) EC(p)   TFML(t)

Tuesday 22/11/94 (Day 2)
========================
 9.00 - 10.30:  CBR(t)  ILDB(t) FLFC(t) KAM(t)  HS(t)
10.30 - 11.00:  Morning Tea Break
11.00 - 12.30:  CBR(t)  ILDB(t) FLFC(t) KAM(t)  HS(t)
12.30 -  2.00:  Lunch
 2.00 -  3.30:  CBR(t)  ILDB(p) FLFC(p) KAM(d)  HS(d)   
 3.30 -  4.00:  Afternoon Tea Break
 4.00 -  5.30:  CBR(p)                               

--------------------------------------------------------------------------------
Examples of structured sequence of tutorials:
--------------------------------------------------------------------------------

There are couple of structured sequence of tutorials that one could adopt.
For example, if the participant is interested in logic/theoretical basis of
AI, then he/she may want to select the following sequence:

Day 1   9.30 - 10.30: Introdution to AI
       11.00 - 12.30: NMR(t)
        2.00 -  3.30: NMR(t)
        4.00 -  5.30: NMR(p)

Day 2   9.00 - 10.30: CBR(t)    
       11.00 - 12.30: CBR(t)
        2.00 -  3.30: CBR(t) 
        4.00 -  5.30: CBR(p)

Alternatively, if the participant is more interested in the applications of
AI, then the following sequence may be more suitable:

Day 1   9.30 - 10.30: Introdution to AI
       11.00 - 12.30: IDSS(t)
        2.00 -  3.30: IDSS(t)
        4.00 -  5.30: IDSS(p)

Day 2   9.00 - 10.30: FLFC(t)
       11.00 - 12.30: FLFC(t)
        2.00 -  3.30: FLFC(p) 


If the interest is in Machine Learning/Knowledge Acquisition, then the
possible sequence may be:

Day 1   9.30 - 10.30: Introdution to AI 
       11.00 - 12.30: TFML(t)
        2.00 -  3.30: TFML(t)
        4.00 -  5.30: TFML(t)

Day 2   9.00 - 10.30: ILDB(t)  or  KAM(t)    
       11.00 - 12.30: ILDB(t)  or  KAM(t)
        2.00 -  3.30: ILDB(p)  or  KAM(d)

Another possible sequence may be:

Day 1   9.30 - 10.30: Introdution to AI
       11.00 - 12.30: EC(t)
        2.00 -  3.30: EC(t)
        4.00 -  5.30: EC(p)

Day 2   9.00 - 10.30: HS(t)    
       11.00 - 12.30: HS(t)
        2.00 -  3.30: HS(d)
--------------------------------------------------------------------------------
Further Information:
--------------------------------------------------------------------------------
For further information on the structured sequence tutorials, please
contact the AI'94 Tutorial Co-ordinator, at the following address:

                        Dr. Dickson Lukose
         Department of Mathematics, Statistics, and Computing Science
                    University of New England
                        Armidale, N.S.W., 2351
                             AUSTRALIA

                 e-mail: ai94@fermat.une.edu.au
                 fax.  : (+61 67) 73 3312
--------------------------------------------------------------------------------