Re: Optimal Prediction

From: Bill Jefferys <>
Date: Thu, 28 Mar 2002 13:27:18 -0500

At 9:32 AM -0800 3/28/02, Wei Dai wrote:

>Perhaps you're not familiar with the history of this mailing list, but
>Juergen Schmidhuber is one of the first authors to explicitly state the
>idea that all possible universes exist in a published scientific paper,
>and that paper is cited in the public invitation to this mailing list.

Good for him. It's an interesting idea.

>have to say it's very inappropriate to publicly "plonk" him when it seems
>to be you that doesn't understand the issue at hand.

I believe that I do understand the issue at hand. It seems that
Juergen and I are using "predictive power" in different senses.

>I agree with Juergen that AP does not predict anything by itself. You need
>at least a probability distribution or measure of some sort in addition to
>AP. There has been extensive discussion on this list over exactly what
>measure should be used, and what the measure is over.

I think that if this is the issue then we are looking at different
interpretations of "predictive power", as I said. My POV is Bayesian.
I am not thinking in terms of what you can say about ensembles of
universes, but about what you can predict about *this* universe,
given the background information that we are here, and other
background information that we may already have (about physics for
example). One can then write down things like P(a certain nuclear
energy level is between x and y|we are here & nuclear physics works
in the way we have learned that it does)=1, which is a prediction of
where we will observe that nuclear energy level to lie when we get
around to observing it (which hadn't been done when Hoyle wrote his
paper). That in my book counts as a real prediction of physics, and
it is a consequence of the AP...and of other things, I agree. If
Juergen's point is that you may need other information than just the
AP (such as how particles will behave in certain physical
situations), then I will stipulate that. But that's a trivial point.
All theories require background information, and a Bayesian is
required to include all background information when making inferences

For me (and I believe most scientists) a theory has predictive power
if it allows us to accurately predict verifiable facts about *this*
physical world, facts that we did not already know. That's exactly
what Hoyle did, and it involved the AP, since (as Barrow and Tipler
point out) we would not exist (certainly in our present form) if
there were no carbon.

I will point out again, what you have not addressed, that people like
Barrow and Tipler, and Tegmark, seem to regard the carbon prediction
as a prediction of the AP.

>Have you read Juergen's paper "A Computer Scientist's View of Life, the
>Universe, and Everything"? If not please do so before making any more
>posts. It's at

Yes, I have read it. He's talking about ensembles of universes, which
isn't what I am talking about. I don't really disagree with anything
I saw in that paper, but it's a different topic from what I was

Since Juergen and I seem to be talking at cross-purposes about
different sorts of predictions, I will put him back on my list to
read. I apologize to him for getting annoyed. But I wish that he
would agree that there is predictive power in the AP as *I* am using
the term, even if not as *he* is using the term. In other words, that
his original statement was painting with too broad a brush.

The AP does have predictive power in the sense that I (as a natural
scientist) usually mean it.

Received on Thu Mar 28 2002 - 10:30:34 PST

This archive was generated by hypermail 2.3.0 : Fri Feb 16 2018 - 13:20:07 PST