Author: Robert Hyatt
Date: 07:06:36 09/13/99
Go up one level in this thread
On September 13, 1999 at 01:22:12, blass uri wrote: >On September 12, 1999 at 20:54:47, Robert Hyatt wrote: > >>On September 12, 1999 at 10:43:37, Randy Schmidt wrote: >> >>>I do not believe that Crafty running on any processor(s) would >>>be stronger than Hiarcs7.1, or for that matter Junior. My >>>large caveat is that the time control be something like eight >>>hours a move (perhaps even 50 hours a move). >>> >>>My point is that the positional elements of Junior and Hiarcs >>>would supercede the speed of crafty on a souped up computer. >>>On any time control faster than 40/2, I think Crafty would have >>>a decisive advantage. >>> >> >> >>Here's a point to ponder. If "junior" has a lot more 'positional understanding' >>than crafty, how would you explain the fact that it is _far_ faster than crafty. >>In fact, it is likely the fastest program running that I have seen NPS numbers >>for. > >The theory that slow searchers are better positional understanding is not a >right theory because the question is not nps for second but if the evaluation >function is good. > >For example I think that Crafty has better positional understanding relative to >tal because tal is too optimistic about the positional advantage. > >It is not clear to me if Junior is a better positional player relative to crafty >but you cannot learn about it by the number of nodes per second. > >I believe that the latest version of Junior is better in >positional understnding relative to previous versions and it is looking at the >same number of nodes per second. > >Uri I agree partially... but there _is_ a direct correlation between NPS and "amount of stuff" in the evaluation. IE in my code, the evaluation is about 50% of the total search time. In Hiarcs I would bet it is closer to 75-80%. In Junior I'd guess at 10%-20% max. Does that mean I do more? Probably. Does that mean mine does better? Not necessarily. Tuning also plays an important role, of course. But one thing is pretty clear. you can't go fast _and_ do a thorough eval. You have to depend more on piece/square stuff and quick things you can detect. And you run into trouble in the right kinds of positions... like when you don't handle outside passed pawns against a program that does, you get ripped by that repeatedly. Or where you don't understand something like Bxh2 or Bxa2, and you get ripped. And I won't name names about the programs that _still_ fall for this one... but the ones that are very fast do, the ones that are slower don't. The reason is pretty obvious... :)
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.