Probabilistic Reasoning
Probabilistic reasoning as information compression by multiple alignment,
unification and search: an introduction and overview
Journal of Universal Computer Science 5(7), 418-462, 1999 (from
2007-01-01, this journal will be an open-access journal and all papers,
including this one, may be downloaded without any charge),
PDF,
uk.arxiv.org/abs/cs.AI/0307010.
This article is a short version of the report, below, and of the
three articles below the report.
Probabilistic reasoning as information compression by multiple alignment, unification and search
SEECS Report, December 1998.
This report is not huge but readers (or potential readers) may wish to know that it
contains 144 pages: 121 pages of text with some diagrams in the 'report with figures 1, 26 and 41'
plus an additional 23 pages of figures in the 'remainder of the figures'.
Postscript: (Report with figures 1, 26 and 41,
Remainder of the figures), Compressed Postscript.
This report describes how a range of constructs and phenomena related to
probabilistic reasoning can be modelled in the SP framework
(see below). These include: probabilistic and fuzzy pattern
recognition and information retrieval; 'indirection' in information
retrieval; one-step reasoning; the integration of 'deductive' and
'abductive' reasoning within a single framework; inheritance of
attributes in a class hierarchy or heterarchy; chains of reasoning
(including decision networks, decision trees, and the operation of 'rules'
in expert systems); the modelling of Bayesian networks (with an example
of 'explaining away'); geometric analogy problems; reasoning with
default values; and nonmonotonic reasoning.
Immediately preceding this report is a short version in the form of
a published paper and, below, are a conference paper and three
unpublished articles based on the report but with some revisions.
Information compression by multiple alignment, unification and search
as a framework for human-like reasoning
Paper presented at FAPR2000: International Conference on Pure and
Applied Practical Reasoning, Imperial College London, 18-20 September 2000. Published in the Proceedings (Imperial College London Department of Computing Report, ISSN 1469-4166, eds. Jim Cunningham and Dov Gabbay, September 2000, pp. 21-36). Now published in the
Logic Journal of the IGPL,
Vol. 9, Issue 2, March 2001, pp. 205-222.
Postcript, Compressed Postscript.
This is essentially a short version of the paper above, with different examples.
Probabilistic reasoning as information compression by multiple alignment,
unification and search (I): Introduction
SEECS, January, 1999.
Postscript: (Article, Figure 1,
Figures 4 and 5). Compressed Postscript.
This article introduces the SP framework, describes the SP61 model
with simple examples of what it can do and introduces the idea that
probabilistic inferences may be drawn from alignments.
Probabilistic reasoning as ICMAUS (II): calculation of probabilities,
best-match pattern recognition and information retrieval,
and reasoning with networks, trees and rules
SEECS, January, 1999.
Postscript: (Article, additional figures).
Compressed Postscript.
This article continues the theme of the previous article. The method of calculating the probabilities of inferences is described. Examples are
presented showing how the ICMAUS framework can model best-match pattern
recognition and information retrieval, including medical diagnosis and
the recognition of objects in a class hierarchy with inheritance of
attributes. Examples are also given of how the ICMAUS framework can
model reasoning with discrimination nets, discrimination trees and
rules of the kind used in expert systems.
Probabilistic reasoning as ICMAUS (III): hypothetical reasoning,
geometric analogies, default values, nonmonotonic reasoning, and
modelling 'explaining away'
SEECS, January, 1999.
Postscript: (Article, Figure 12,
additional figures). Compressed Postscript.
This article continues the theme of the previous two articles. Examples
are presented showing how the ICMAUS framework can accommodate hypothetical
("what if") reasoning, solving geometric analogy problems, reasoning with
default values and nonmonotonic reasoning. The framework provides an
alternative to the Bayesian network explanation of 'explaining away'.
|