Author: Robert Hyatt
Date: 18:35:23 10/07/02
Go up one level in this thread
On October 07, 2002 at 19:40:16, Tom Kerrigan wrote: >1. Deep Blue _did_ do forward pruning. They used null move for threat detection, >not forward pruning, but they did have "domain specific" forward pruning. Hsu >emphasized this quite a bit. He said specifically that it was responsible for >allowing them to search critical variations very deeply. > >2. Deep Blue did _not_ have impossibly sophisticated eval terms. Hsu gave two >examples of DB's "extremely costly" eval terms: > > i. Seeing if an open file is valuable by calculating the square control of >the 7th and 8th ranks of the file (including rays) > ii. Doing king safety calculations for both sides of the board while the >king is uncastled. > >Many PC programs calculate square control and could easily/cheaply implement >(i), and (ii) is not THAT expensive. If king safety takes 10% of a PC program's >time, doing it twice would only make it 10% slower (and only in positions where >kings aren't castled). > >Moreover, Hsu said that a major limitation of DB1 (that was fixed in DB2) was >that square control information could only be used for a limited number of eval >terms, apparently because it was available late in the eval pipeline. So there >were limitations to the eval terms that could be computed by DB. A couple of points. (1) nobody said they had "impossibly complex eval terms". I (and the DB team) _did_ say that they could do whatever they wanted. If something is ready "too late" all you need to do is extend the pipeline a cycle, and let -er rip.... _if_ you really want that eval term and it is serially based on something that has to be computed first... Compare that to what _we_ have to do which is to serially compute _everything_. (2) what is so important about DB1? DB2 was _the_ machine... and he fixed a _lot_ of problems from DB1 in the re-design... > >3. There _was_ an implementation of DB's algorithms in software. I asked how the >new eval terms in the DB2 chips were tested & tuned without actually having the >chips, and Hsu said that Joel Benjamin played a "software simulation" to test >them. He made it sound like that should have been obvious. (Are you listening, >Bob?) Never said there was _not_ one. I said "there was no reasonably efficient software implementation" that could be used in an engine." Big difference. You _can_ emulate _anything_ in software, by definition. But that doesn't mean you can use it to play games... They discussed their "software tuning" more than once, and apparently had some sort of GUI interface to the process... > >4. While he does have the rights to the DB chip design, Hsu is still bound by >IBM NDA for the eval terms and is not legally allowed to disclose them. > Nothing surprising there. >-Tom
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.