Author: Christophe Theron
Date: 09:10:26 07/04/02
Go up one level in this thread
On July 04, 2002 at 10:07:37, Robert Hyatt wrote:
>On July 04, 2002 at 03:49:40, Uri Blass wrote:
>
>>On July 03, 2002 at 14:29:17, Robert Hyatt wrote:
>>
>>>On July 03, 2002 at 13:46:17, Christophe Theron wrote:
>>>
>>>>On July 02, 2002 at 20:20:37, Robert Hyatt wrote:
>>>>
>>>>>On July 02, 2002 at 18:54:49, Keith Evans wrote:
>>>>>
>>>>>>
>>>>>>Sorry to be anal retentive, but that's a bit of a stretch. Here's what they say:
>>>>>>
>>>>>>"The chess chips optionally support the use of an external FPGA (Field
>>>>>>Programmable Gate Array) to provide access to an external transposition table,
>>>>>>more complicated search control, and additional terms for the evaluation
>>>>>>function. In theory this mechanism would have allowed the hardware search to
>>>>>>approach the efficiency and complexity of the software search. Null move search
>>>>>>was also explicitly supported by this method. Due to time constraints, this
>>>>>>capability was never used in Deep Blue."
>>>>>>
>>>>>
>>>>>
>>>>>Read on. On page 67, section 4.1, item 3, "mate threat".
>>>>>
>>>>>"It is relatively simple using a null move search to detect if there is a
>>>>>threat in the current position.... The Deep Blue implementation ...
>>>>>
>>>>>Which matches what I said. They had support for a normal null-move search
>>>>>had they wanted to use it, but they did use null-move to detect threats,
>>>>>something that has been done before (and several of us use a form of mate
>>>>>threat extension based on this idea presently).
>>>>>
>>>>>So they used null-move in at least one way, without using it as a forward
>>>>>pruning algorithm, which fits with Hsu's "no errors in the search" theme he
>>>>>mentioned repeatedly over the years. Extra extensions were one thing to him,
>>>>>but outright errors were something else not to be tolerated. Right or wrong.
>>>>>I obviously disagree about the errors in a normal null-move search, but I
>>>>>can hardly argue with their success...
>>>>
>>>>
>>>>
>>>>That's my point as well.
>>>>
>>>>I don't argue about their success.
>>>>
>>>>I'm just saying that they succeeded because their chips were very fast. So fast
>>>>that they allowed them to use inferior search techniques and still succeed.
>>>>
>>>
>>>Could you not make the _same_ statement about chess 4.0 in 1975? Until that
>>>point _everybody_ was doing forward pruning like mad. They discovered that a
>>>a shallower full-width search had fewer errors and they stomped everybody into
>>>the ground until everyone converted...
>>
>>It is different.
>>It is obvious that selective search from the first plies
>>is a mistake when you have speed.
>>
>>It also seems obvious that pruning rules that are based
>>on the remaining depth is a good idea and you can use them
>>and see everything if you search deep enough.
>>
>>Uri
>
>
>Everybody is overlooking an _important_ detail, so lets take this back to
>CS101:
>
>1. Forward pruning is a form of selective search. You cull moves you think
>are no good, so that the rest are basically "extended" or searched deeper than
>the "lemon" moves.
>
>2. Search extensions do _exactly_ the same thing. They extend the moves you
>think are "good" so that they are searched more deeply, while the ones you
>do not extend are not searched that deep.
>
>In simple terms, the two ideas are _identical_ in every way, as far as the
>final result. To say that doing a full-width search with lots of very
>sophisticated extensions is not as good as doing a sophisticated selective
>search (forward pruning) is not a particularly sensible statement to make.
>
>_anybody_ that has spent any time on tree-searching will realize that _either_
>will produce _exactly_ the same result assuming the extensions and forward-
>pruning are done with the same skill level.
>
>So picking on this aspect of deep blue is simply a strawman argument. They
>clearly do more extensions than the rest of us. Which _may_ offset their
>lack of forward pruning. Believing or claiming anything else shows a lack
>of understanding of something...
In this case, claiming that you are doing brute force just because you do not
want errors in your search is also a lack of understanding.
Didn't Hsu say this? Aren't you repeating his words every time you can?
First your point is that they have picked brute force because they had enough
power and did not want mistakes in the search, and now you are saying that they
had a selective search and that it is equivalent to what can be achieved with
strong pruning.
Thinking about it, it seems that you can indeed get the same search enveloppe by
either pruning or extending.
But thinking twice about it, I think that it is not possible with the search
extension techniques used in Deep Blue to get something equivalent to the simple
and efficient pruning techniques we know today.
I smell that there is something important behind this and I will have to think
more about it. That's an interesting research area.
>As far as your selective search comments, It is obvious (to me) that everybody
>is not doing selectivity just deeply in the tree. It is being done near the
>root as well, based on some very trivial oversights that some programs make from
>time to time. Oversights that a 4 ply full-width search would see.
It's not as simple as that.
"Near the root" can mean several different things.
You can apply some kind of gross pruning system near the root and make big
shortsighted mistakes.
You can also apply some detection near the root and collect information to prune
later. Then you don't make such big mistakes.
The argument that pruning will make obvious blunders sometimes is simply wrong.
Christophe
This page took 0.02 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.