Account Options

  1. Sign in
The old Google Groups will be going away soon.
Switch to the new Google Groups.
Google Groups Home
« Groups Home
comp.compression
+ new post
About this group
Subscribe to this group
This is a Usenet group - learn more
Message from discussion The C-Prize
The group you are posting to is a Usenet group. Messages posted to this group will make your email address visible to anyone on the Internet.
Your reply message has not been sent.
Your post was successful
 
From:
To:
Cc:
Followup To:
Add Cc | Add Followup-to | Edit Subject
Subject:
Validation:
For verification purposes please type the characters you see in the picture below or the numbers you hear by clicking the accessibility icon. Listen and type the numbers you hear
 
Matt Mahoney  
View profile  
 More options Jun 19 2005, 2:50 pm
Newsgroups: comp.compression
From: "Matt Mahoney" <matmaho...@yahoo.com>
Date: 19 Jun 2005 11:50:37 -0700
Local: Sun, Jun 19 2005 2:50 pm
Subject: Re: The C-Prize
Print | View thread | Show original | Report this message | Find messages by this author

jim_bow...@hotmail.com wrote:
> Matt Mahoney writes:
> > He proves... the most likely outcome for
> > any experiment is the one with the simplest explanation, where
> > "simplest" means the smallest program that could model what you
> > currently know about the universe.

> Wow.

> If what you say is easily verifiable by consulting mathematicians then
> you've just dramatically strengthened the case for the C-Prize.  A
> philanthropist can have a trusted consultant verify Hutter's proof and
> the C-Prize's salience to it.  The only remaining question is whether
> the C-Prize is properly administered -- something the philanthropist
> can verify with a combination of business and engineering consultants.

I don't think there is much problem with the proof so much as there is
with arguing whether it means anything.  Hutter gives a formal
definition of an agent, but then we have to make the leap that an agent
is a good model of human behavior.  (It seems so).  Another assumption
is that the universe is computable.  Again it seems like a reasonable
assumption, but others might not agree.

> I had run across Hutter's AIXI paper before, when searching for
> Solomonoff's algorithmic probability theory, but I read just the
> abstract and didn't have the background for it so I sent it off to a
> friend who is a doctoral candidate in statistical mechanics to review.
> Being a doctoral candidate he doesn't have much time left to review
> papers ourside his area so I never got the low-down you just gave.
> Thanks for the clear synopsis.

> A proof like this should have been available to researchers from the
> 1960s work on minimum length description (AP, KC, etc.) and decision
> theory but apparently no one bothered to put 2 and 2 together for what
> -- 40 years now?

> This is going to be a major question for science historians:

> Why did the AI community fail to formalize this proof for so long?

> Given Moore's Law originated at about that time, 40 years ago, there's
> a good argument that this ranks as one of the greatest losses of
> potential to befall humanity.

I think the AI community ignored it is because it does not lead to a
practical solution to AI.  Finding the shortest program consistent with
some data is not computable.  Hutter describes a restricted version
which runs in exponential time, but again this is not very useful.
What it does is formalize the goal of machine learning to generalize to
unseen data, something researchers have already been doing.  Hutter's
work puts this in a consistent framework.

-- Matt Mahoney


 
You must Sign in before you can post messages.
To post a message you must first join this group.
Please update your nickname on the subscription settings page before posting.
You do not have the permission required to post.
Create a group - Google Groups - Google Home - Terms of Service - Privacy Policy
©2012 Google