Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: next deep blue

Author: Jeremiah Penery

Date: 02:37:21 01/24/00

Go up one level in this thread


On January 23, 2000 at 19:40:47, Tom Kerrigan wrote:

>On January 23, 2000 at 19:29:31, Jeremiah Penery wrote:
>
>>>No, the only reason you think DB has a bunch more evaluation is because Hyatt
>>It has nothing to do with Hyatt.  Please try to construct better arguments in
>>the future.  The reason I think DB has a better evaluation is that I've seen the
>>games, and analyzed them.  Kasparov and other GMs have said that DB was clearly
>>superior to anything else they've seen.
>
>Of course it's superior, it searched 200M NPS. Searching has the property of
>"adding knowledge" to a program. How do you know that you were seeing evaluation
>function terms in those games, and not tactics that are so deep that they're
>hidden to humans?

So why didn't he think this of DB-1?  It did about 100M NPS, but he ended up
crushing it (see game 6 of the first match).  The NPS didn't seem to help so
much there.  Obviously, there was a lot more knowledge in DB-2.

>>>keeps saying that it does. But none of us has any proof that DB has more
>>>evaluation than CS Tal. And Hsu doesn't even think it has more evaluation,
>>>because his estimate translates to at least 20k NPS.
>>That was an estimate.  He could have very well been way off, though it's not
>>that likely.

Now that I think about it more, it _is_ likely.  See below.

>Right. If anything, I suspect the estimate is high. Hsu made the estimate to
>compare the DB chip to general-purpose computers. The higher the estimate, the
>more impressive his work sounds. I'm not saying that Hsu is a liar, or
>misestimated on purpose, but I think it's unlikely that Hsu erred on the side of
>DB chips doing LESS work.

Take some generally-used evaluation terms.  How many instructions does it take
to apply them, on average?  At least a few?  At worse a few hundred?

Now take some not-so-generally-used terms. E.g., which pieces attack squares
around the king, including attacks through any other pieces and
batteries/x-rays; potentially open files; anything else they were doing that
might be expensive.  How many instructions for these?  A few thousand, maybe?
More?

With some 8k terms in their evaluation, each term would have to be carried out
in an average of 5 instructions to equal Hsu's 40k.  How much can you do in 5
instructions?  Not much.
It's likely he erred on the side of DB doing less work by a large margin.



This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.