Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!uflorida!indri!aplcen!jhunix!ins_atge
From: ins_atge@jhunix.HCF.JHU.EDU (Thomas G Edwards)
Newsgroups: comp.ai
Subject: Re: free will and errant gamma rays
Summary: Random events in Connectionism
Message-ID: <1528@jhunix.HCF.JHU.EDU>
Date: 21 Apr 89 19:54:16 GMT
References: <1499@hub.ucsb.edu>
Reply-To: ins_atge@jhunix.UUCP (Thomas G Edwards)
Organization: The Johns Hopkins University - HCF
Lines: 21

In article <1499@hub.ucsb.edu> silber@sbphy.ucsb.edu writes:
>So as far as man made a-i machines are concerned, maybe, at some point,
>one might want to specifically include some physical random event
>generator (a lump of radium or some such radioactivbe source) ?????????

I'd say that there are plenty of AI paradigms making great use of
random number generators!  When you train a neural-net, you usually
begin with small random weights between the neural units to break the
symetry of the network so the gradient descent process can change
every weight individually.

Boltzman machines (a kind of connectionist learning routine which
involes "simulated annealing," or slowly letting the network relax
into a low energy state by changing how fast a neural element can change
state) depend on random numbers, since neural elements have probabilities
of being on.

In genetic algorithms, random number generators are essential to
adding genetic novelty to the search space of the algorithms.

-Thomas Edwards