RESEARCH INTO THE CONSTRUCTION AND APPLICATION OF INTELLIGENT KNOWLEDGE
                  BASED SYSTEMS THAT AID COMPUTER USERS

                            Dr. Peter Ross,
                  Department of Artificial Intelligence,
                        University of Edinburgh.
   
INTRODUCTION

   Intelligent knowledge-based (IKB) systems, sometimes referred to as 'expert
systems', are computer programs or sets of programs which contain explicit
knowledge concerning some particular area of human expertise, organised in
such a way that they can function as an expert consultant in that area. Some
existing examples are MECHO (which solves problems in mechanics - [Bundy79]),
MYCIN (which diagnoses bacterial infections - [Shortliffe76]), MOLGEN (which
designs molecular genetic experiments - [Stefik81]) and EL (which analyses
analogue electric circuits [Stallman77]). They contain the factual knowledge
of their subject as sets of rules - for example, one of MYCIN's rules is

   If (a) the infection is primary-bacteremia, and
      (b) the site of the culture is one of the sterile sites, and
      (c) the suspected entry point is the gastro-intestinal tract,
   Then there is suggestive evidence (prob. 0.7) that the organism is
      bacteroides.

MYCIN functions by collecting evidence, through questioning a physician, to
formulate various hypotheses and weighing the evidence for and against them
using rules such as the one above. These rules are also used to guide the
questioning, for instance to establish whether or not the preconditions of 
certain rules are probably or certainly true or false.

   Much of the current research into intelligent knowledge-based systems 
is concerned with two questions. The first is, how does one represent and 
structure the various necessary sorts of knowledge so as to be usable by a 
computer program? The second question is, how can a program use the various
sorts of knowledge and yet avoid a combinatorial explosion of the number of
possibilities that need to be considered? Certain useful principles are 
already emerging (e.g. see [Hayes-Roth82]), but usually the solutions are 
tailored to fit the area of knowledge in question. 

   One option that is open to almost all existing IKB systems is to ask the
user some pertinent questions when there is difficulty deciding what to
pursue next, on the assumption that all the user lacks is the expertise to 
make the best use of his own information. Little work has been done on the 
problems that arise when the user is using the IKB system to acquire expertise 
himself, and therefore may not be able to understand the questions or may not 
be able to formulate the answers properly. In such a situation the system 
would have to form a model of what the user knows and can do at his current 
level of competence, and use the model to find the right level at which to 
communicate with the user. Such a model would have to include not only details 
of his factual knowledge but also details of his higher-level knowledge of how
to apply that knowledge. There are probably two resons why not much work has
yet been done in this area. One is that the system would have to begin by
being a relatively passive observer of the user's efforts, since it would be
unable to assess whether particular questions would be meaningful to him;
however there are major linguistic difficulties in extracting the meaning
from a user's utterances if he is allowed to use even a small subset of
English. The other is that in almost all subject areas the vocabulary needed 
by the system to represent the higher-level knowledge is too large for the
problem to be tractable.

PROPOSED RESEARCH

   The research I wish to do is to investigate the problem of an IKB system 
forming and using a model of the user's knowledge and competence. To avoid the 
difficulties mentioned above, I propose to use as subject area the topic of 
how to use a computer operating system, in particular the UNIX system now 
coming into widespread use. This restricts the user's utterances to the range 
of commands meaningful to the operating system and these have well-defined 
syntax and semantics, neither impossibly large. Moreover, the function of an 
operating system is to allow users to accomplish certain kinds of tasks by 
making certain kinds of operations easy to perform. This set of operations 
provides a natural vocabulary for representing the user's intentions, and it 
is a vocabulary which the user himself has to acquire in learning how to use 
the computer.

   The work breaks naturally into several tasks:

 (a) constructing a database of static knowledge of the system, such as 
     information about the available commands, what parameters they take, what 
     they do and what effects they produce
 (b) constructing an IKB program that has knowledge of the use of the system.
     It will include rules such as
        IF his program failed THEN the debris is in a file called 'core'
        IF he couldn't delete a file whe he tried AND it exists but is
           protected against deletion AND there is a low probability that
           he knows about the file security mechanism THEN tell him about it
           and increase the probability that he knows about the mechanism
     and many others
 (c) constructing an IKB program that has general rules about any users'
     possible strategies and intentions, e.g. "good programmers normally keep
     large programs in several files rather than one unwieldy one", "good
     programmers normally name files in a logical and consistent fashion"
     and rules that express the many different kinds of program development 
     cycle that users might choose to adopt
 (d) constructing an IKB program that can draw on the knowledge and rules
     described above, and use inferencing and pattern-recognition mechanisms
     to attempt to model the user's knowledge, aims and competence. Some of 
     these mechanisms will be for essentially mundane tasks such as 
     recognising groups of files that have indeed been named in a logical and
     consistent manner, others will be for the key tasks of forming hypotheses
     about the user and then looking for evidence that confirms or refutes 
     them.

   Volunteers from the (large) community of computer users in the Department 
of Artificial Intelligence, some of whom are highly experienced and some of 
whom are novices, will be asked to use the programs so that I can assess 
progress and get useful fundamental information for directing the work.

   The most suitable tool for the work is PROLOG, a very high level computer
language based on a particular formulation of predicate logic. The Department
of Artificial Intelligence wrote and now distribute a UNIX PROLOG system;
moreover, the MECHO program mentioned above was written in PROLOG by members
of the department led by Dr. Alan Bundy, and it should be possible to draw on
their experience when tackling tasks (a), (b) and (c).

BENEFITS OF THE RESEARCH

   In addition to furthering research into IKB systems, the proposed work 
should also be useful to those working on problems of cognitive modelling and 
of man-machine communications. It should also provide a practical tool for 
those people needing to learn to use UNIX, and make interesting contributions 
to the subject of computer-assisted instruction.

RELATED RESEARCH

   Apart from research into IKB systems at the Departments of Artificial 
Intelligence and machine intelligence at Edinburgh University and a few other 
centres, most of the work on this subject area in the U.K. is directed toward
modifying the operating system itself to make it more friendly and tolerant of 
novices. Probably the best system that creates a model of the user has been 
one called WEST [Brown78] written in the U.S.A., which coaches children in 
simple arithmetic skills by monitoring how they play a simple 'snakes and 
ladders' type of game in which the child moves an amount determined by 
combining three random numbers in any simple arithmetic way. If nothing else, 
WEST indicates that the proposed research is, at least to a certain extent, 
possible using current ideas in Artificial Intelligence.