| Introduction
In the late 1980's Knowledge Based Systems (KBS)
were seen to be leading edge software technology. Developers thought that the
simplest KBS paradigm, Expert Systems, perhaps combined with probabilistic and
fuzzy logic extensions would soon revolutionise the way that software was used
throughout business and other sectors of the economy.
KBS
software was built on rules which encoded the knowledge of experts in any given
domain. Computers would then use this encoded knowledge to make decisions on behalf
of their human users.
It was not long however,
before the bubble of hype surrounding these systems began to burst. Something
was wrong, but what was it? The Knowledge Acquisition Bottleneck
Apart from the limited power of the computers available
at the time, the major problem was the difficulty of acquiring implicit knowledge
from the minds of experts and then representing it explicitly. This so-called
Knowledge
| | Acquisition
Bottleneck was believed to be the limiting factor on building systems that could
do complex, useful tasks.
By the end of the
twentieth century however, university departments were working hard at this problem.
Curiously it was often Psychology departments rather than Computer Science departments
which had the most impact in this area.
In
particular, Ethnography (by then seen as a core part of Cognitive Psychology)
was being used to study behaviour in situ with the aim of identifying the cognitive
processes underlying that behaviour. Just as Margaret Mead (an early ethnographer)
had lived amongst native tribes in Papua New Guinea in order to study their cognitive
behaviour, so Psychology departments were sending researchers (often under cover)
into workplace environments to discover how people approached problem-solving
activities.
This work was, and continues to
be, very successful. Knowledge acquisition is no longer the 'black art' it was
deemed to be. Despite this, KBS has continued to be underused. Why might this
be? A Knowledge Representation Bottleneck?
It is my contention that the problem was not primarily
with how we obtained knowledge, but with how we represented it. I am not arguing
that rules (or Bayesian networks and other knowledge representation methods) are
inadequate to the |
