@Article{Levesque1986,
author = {Hector J. Levesque},
title = {Making Believers out of Computers},
journal = {Artificial Intelligence},
year = {1986},
volume = {30},
issue = {1},
pages = {81-108}
}
The idea behind KBS is to construct systems that exhibit knowledge which is represented in their data structures. The system then manipulates this knowledge explicitly. An expert level performance can be achieved by applying these principals as well as domain-specific knowledge. In this paper only theoretical aspects of KBS are discussed.
Computers and Thought- According to Leibniz propositional expressions
could be combined using the rules of logic and reach conclusions in mechanical ways. So, if there is a set of prepositions and a conclusion, we are only
concerned with the form. The problem is making computers execute this type of thinking. [p.82-83]
The correct answer can be achieved relying solely on syntactic manipulations. This means being able to describe what the world would be like
if a collection of sentences were true. Or, in more scientific terms, the logical implication of an array of sentences.
In this context logic is a
scientific study of implications, the ability to extract implicit information. Rational inference is not considered as implicit. For example, if
we know Pat is a computer scientist we cannot infer that Pat is a male. [p.85]
The KBSs concerned reason over an explicit knowledge base. These data structures have two characteristics: first, the sentences are part of the knoweldge the
program exhibits; and second, they influence the behavior of the program. [p.86]
A Difficult Case- As the number of cases grows, the time it would take to reach a conclusion grows enormously and results in combinatorial
explosions. Knowledge engineering is not always helpful, since some notions are not domain specific, like sitting constraints at a party. Massive parallelism
is of no help either, because while it provides constant factor speed up, some problems grow exponentially.
Essential to look through all
possibilities as the difficulty seems to be a result of the particular information-processing problem. [p.86-88]
Why are worst case scenarios important? It is hard to decide what an average case or sentence is. Creating a new information-processing problem will merely
postpone the rise of a worst-case problem. We want these programs to be well behaved, such that they will produce the right answer at the right time. We also
want to be able to apply knowledge without requiring more knowledge. Otherwise, the explanation will get into an infinite regress.
A Much Easier Case- Nowadays, knowledge-based AI requires large amounts of domain-dependent knowledge which produce expert systems.
Why is it easy to apply large amounts of domain-dependent information? The kinds of knowledge are limited, so the search space is limited. A small change in the form
of a problem can make a problem un-doable.[p.90-91]
A KB is incomplete if it tells that a sentence it true, but doesn't specify which. Some KB can be complete but still require extended processing. [p.92,93]
Break down complex questions and work in parallel.
A vivid KB- 1:1 correspondence between a class of symbols in the KB and objects in the real world. And, for every relationship in the world, there will be a connection between
symbols in the KB. Vivid doesn't mean precise, as in 'Henry is related to Bill'. Can use direct calculations to determine what is true in the domain of discourse.
Provided ultimate answers to questions. [p.93-97]
"Thinking may sometimes feel like much more than symbol manipulation, but this could be an illusion caused by the very close correspondence between symbols in a vivid KB
and what they are about." [p.95]
A holistic approach to reasoning can be achieved when knowledge is complete and vivid. [p.96]
Pictorial information is limited in form. [p.98, 101]
Logically unsound- the truth of the conclusions is not implicit in the original sentence. We infer existence and non-existence of objects. [p.101]
"Make incomplete knowledge more vivid by filling in details by
using defaults and closed-world assumptions derived from knowledge about
knowledge." [p.102]
Logically incomplete thinking can be transformed into vivid thinking by, for example, converting negation into an atomic
tence. [p.104, 106]
Cognitive Science Summaries Webmaster: