文档介绍：Probabilistic Inference Using
Mark o v Chain Mon te Carlo Metho ds
Radford M. Neal
T ec hnical Rep ort CR G-TR-93-1
Departmen t puter Science
Univ ersit yofT oron to
E-mail: ******@
25 Septem b er 1993
c
Cop yrigh t 1993 b y Radford M. Neal
Abstract
Probabilistic inference is an attractiv e approac h to uncertain reasoning and em-
pirical learning in arti cial in telligence. Computational diculties arise, ho w ev er,
b ecause probabilistic mo dels with the necessary realism and
exibilit y lead -
plex distributions o v er high-dimensional spaces.
Related problems in other elds ha v e b een tac kled using Mon te Carlo metho ds based
on sampling using Mark o vc hains, pro viding a ric h arra yoftec hniques that can b e
applied to problems in arti cial in telligence. The \Metrop olis algorithm" has b een
used to solv e dicult problems in statistical ph ysics for o v er fort yy ears, and, in the
last few y ears, the related metho d of \Gibbs sampling" has b een applied to problems
of statistical inference. Concurren tly , an alternativ e metho d for solving problems
in statistical ph ysics b y means of dynamical sim ulation has b een dev elop ed as w ell,
and has recen tly b een uni ed with the Metrop olis algorithm to pro duce the \h ybrid
Mon te Carlo" metho d. puter science, Mark o vc hain sampling is the basis
of the heuristic optimization tec hnique of \sim ulated annealing", and has recen tly
b een used in randomized algorithms for appro ximate coun ting of large sets.
In this review, I outline the role of probabilistic inference in arti cial in telligence,
presen t the theory of Mark o vc hains, and describ e v arious Mark o vc hain Mon te
Carlo algorithms, along with a n um b er of supp orting tec hniques. I try to presen ta
comprehensiv e picture of the range of metho ds that ha v e b een dev elop ed, including
tec hniques from the v aried literature that ha v e not y et seen wide application in
arti c