Abstract
In this paper we present an information theoretic measure for the unsupervised inference of Probabilistic Recursive Transition Networks and Probabilistic Finite State Automata from bodies of training data. The method we present here is based on the Minimum Message Length principle. It constructs a compact code or message which contains a trial theory (automaton) about the data followed by the data encoded using the theory. The trial theory with the shortest message length (including the data) contains the most likely explanation theory about the data among all the trial theories. We do not present an algorithm for generating the trial theories as it may vary depending on the domain of the problem.
| Original language | English |
|---|---|
| Pages (from-to) | 395-400 |
| Number of pages | 6 |
| Journal | National Conference Publication - Institution of Engineers, Australia |
| Volume | 1 |
| Issue number | 94 /9 |
| Publication status | Published - 1 Dec 1994 |
| Event | Proceedings of the International Symposium on Information Theory & Its Applications 1994. Part 1 (of 2) - Sydney, Aust Duration: 20 Nov 1994 → 24 Nov 1994 |
Cite this
}
Compact coding of automata for inference. / Collins, Matthew; Zukerman, Ingrid.
In: National Conference Publication - Institution of Engineers, Australia, Vol. 1, No. 94 /9, 01.12.1994, p. 395-400.Research output: Contribution to journal › Conference article › Research › peer-review
TY - JOUR
T1 - Compact coding of automata for inference
AU - Collins, Matthew
AU - Zukerman, Ingrid
PY - 1994/12/1
Y1 - 1994/12/1
N2 - In this paper we present an information theoretic measure for the unsupervised inference of Probabilistic Recursive Transition Networks and Probabilistic Finite State Automata from bodies of training data. The method we present here is based on the Minimum Message Length principle. It constructs a compact code or message which contains a trial theory (automaton) about the data followed by the data encoded using the theory. The trial theory with the shortest message length (including the data) contains the most likely explanation theory about the data among all the trial theories. We do not present an algorithm for generating the trial theories as it may vary depending on the domain of the problem.
AB - In this paper we present an information theoretic measure for the unsupervised inference of Probabilistic Recursive Transition Networks and Probabilistic Finite State Automata from bodies of training data. The method we present here is based on the Minimum Message Length principle. It constructs a compact code or message which contains a trial theory (automaton) about the data followed by the data encoded using the theory. The trial theory with the shortest message length (including the data) contains the most likely explanation theory about the data among all the trial theories. We do not present an algorithm for generating the trial theories as it may vary depending on the domain of the problem.
UR - http://www.scopus.com/inward/record.url?scp=0028744619&partnerID=8YFLogxK
M3 - Conference article
VL - 1
SP - 395
EP - 400
JO - National Conference Publication - Institution of Engineers, Australia
JF - National Conference Publication - Institution of Engineers, Australia
SN - 0313-6922
IS - 94 /9
ER -