Coping with Surprise by sdfsb346f

VIEWS: 4 PAGES: 3

More Info
									CONTRIBUTION to the GRAND CHALENGES meeting, Edinburgh, November 2002.




                                         Coping with Surprise

                                              Derek Sleeman1
                                        Computing Science Department
                                               The University
                                          ABERDEEN AB24 3FX
                                            Phone: +44 (0)1224 272288/96
                                           Email: dsleeman@csd.abdn.ac.uk


The Vision
It appears that young children find out much about their surroundings by creating
hypotheses about how the world works, and then revising these ideas once they get
feedback from their surroundings. For example, a reasonable hypothesis about the
weights of objects might say that larger objects weigh more than smaller objects. It is
generally agreed that such feedback is more effective if it is from real experience (like
touching & feeling) rather than being provided by some authoritative figure, such as a
parent or teacher. Thus at an early stage children develop an empirical approach to
knowledge discovery. (See work by Klahr at CMU & Ross Driver's book on Science
Education.)


Similarly, it can be argued that adults in many aspects of their lives also spend a great
deal of time dealing with surprise. I'll mention the situation with just 2 groups here:
Scientists & Clinical Doctors2.


Classically, in the field of Physics, Newton's laws of motion were held for many decades
to be highly predictive for motion; then, people started taking observations of fast-
moving celestial bodies; at which point a sizeable number of anomalies were detected.


1
    Aberdeen's Advanced Knowledge Technologies PI.
2
 A similar case could be made for managers of complex organizations who make simplifying assumptions
& use these until they are challenged.



\\Umbriel\htdocs\NeSCTest\Copy\Nesc\secure\esi\events\grand_challenges\panelc\sleeman.doc
Finally these anomalies were explained by Einstein who postulated a more
comprehensive model which covered both the observations covered by Newton's models
& the subsequent anomalous data points. Similarly, for a century or more Botanists
accepted a particular taxonomy for a family of plants, Gaultheria; then a new specimen
was identified which caused them to re-access their classification. This actual taxonomic
restructuring was done by Middleton & Wilcock in Botany at Aberdeen; a few years later
Alberdi & Sleeman implemented a system, ReTAX, which was able to closely replicate,
the various steps which the botanists had taken in revising this taxonomy (see our 1997
AIJ paper).


Further, I believe there are huge possibilities in many of the Sciences, including
Medicine, to bridge the gulf between (evolving/partial) theories held by professionals and
the massive data sets which are now available. For example, physiologists have
postulated theories of many organs of the body, including the renal system. But as far as
I am aware no one has tried to use these (partial) theories to predict how patients on
dialysis might respond to treatment (Note modern dialysis machines can provide
extensive data-sets.) The general challenge here then is to seek to reconcile complex
theories with extensive data-sets. In general there are 3 ways in which this might be done.
If one believes the theory then parts of the data can be said to be inconsistent; if one
believes the data then parts of the theory can be questioned. In pratice, as in the Botany
example above, there is the intermediary position where, under the guidance of the
domain expert, both aspects of the data & the theory would be refined/revised to as to
reach a "consistent story". There are of course further challenges here, as in many cases
the partial theories will not be sufficiently detailed to be used predictively. In which
cases, it will be necessary for the domain expert to add assumptions to make the theory,
in some contexts, executable. And again those assumptions will be accepted until they
produce predictions which are unacceptable to the domain expert or to meta-data.




\\Umbriel\htdocs\NeSCTest\Copy\Nesc\secure\esi\events\grand_challenges\panelc\sleeman.doc
Challenges
The challenges here are to build a series of aides which incorporate theories of the world
(of our body organs etc); these aides should support users at a range of levels. Domain
knowledge would be supported by ontologies, and the level of the interaction with the
user would be based on an appropriate user model (young child on the one hand to
consultant physician on the other). Such systems would also reuse domain knowledge
bases in a variety of ways, and so would draw upon the work on knowledge base reuse
which is currently being undertaken by the AKT Consortium. Finally there is also great
potential in such systems for them to enhance their knowledge bases as a result of
feedback from the real world - as did the young child in our opening scenario.




\\Umbriel\htdocs\NeSCTest\Copy\Nesc\secure\esi\events\grand_challenges\panelc\sleeman.doc

								
To top