## Go With the Winners - General Monte Carlo

From: "andrew cooke" <andrew@...>

Date: Mon, 1 May 2006 16:57:24 -0400 (CLT)

This is sweet.
http://www.citebase.org/cgi-bin/citations?id=oai:arXiv.org:cond-mat/0201313

In monte-carlo simulations you waste a lot of time generating common
configurations which give you useless information about areas of the
problem space you already understand.  So throw them away.  And when you
get a configuration probing an unusual (rare) area, clone it.

Now obviously you can't do that.  But if you're generating these
configurations in a sequence of steps then you can do it - as long as you
take care to weight things correctly and make sure discards are unbiased.

Read the paper - near the end of the first page you'll go "ah, yes, of
course!".

From a reference in a book I picked up cheap a while back (whose name
escapes me - the book is in La Serena, I was reading it last night before
getting on the bus to come home).

Andrew