Author: Christophe Theron
Date: 23:14:02 07/06/99
Go up one level in this thread
On July 06, 1999 at 22:20:54, Robert Hyatt wrote:
>On July 06, 1999 at 15:48:35, Christophe Theron wrote:
>
>>On July 06, 1999 at 13:17:48, Robert Hyatt wrote:
>>
>>>On July 04, 1999 at 17:29:35, blass uri wrote:
>>>
>>>>
>>>>On July 04, 1999 at 17:12:02, Bo Persson wrote:
>>>>
>>>><snipped>
>>>>>Not quite.
>>>>>
>>>>>If you run under Windows, a program can behave badly and be a CPU hog. It can do
>>>>>a number of "tricks", like increasing its own priority, to get more CPU time
>>>>>from the system.
>>>>>
>>>>>This will be unfair to "the nice guy" who's program runs "properly" - share and
>>>>>share alike.
>>>>
>>>>I do not suggest thinking and pondering at the same time.
>>>>The only reason that the game is going to be twice longer is that instead of
>>>>thinking and pondering at the same time I suggest to do it not at the same time
>>>>so instead thinking and pondering for 2 minutes on the same time I need 4
>>>>minutes(2 for one engine to think and 2 for the second engine to ponder without
>>>>knowing the move of the first engine)
>>>>
>>>>Uri
>>>
>>>I've explained this several times. "ponder=off" (crafty terminology) is _not_
>>>the way to play engine vs engine games. I do _all_ of my testing with
>>>ponder=on, and only use ponder=off for test suites and debugging. My time
>>>allocation code is tuned to run with ponder=on. Running with it off will
>>>most definitely cause some timing difficulties that are not normally seen.
>>>
>>>I'd bet that if you ask, most programmers test with ponder=on and feel very
>>>comfortable with their code. But if you ask them to play a serious tournament
>>>with ponder=off, I'd bet you would see a _lot_ of testing going on to make sure
>>>that this doesn't break anything.
>>
>>
>>I always test my engine with ponder=off to get (almost) reproducable results,
>>and because this way I need less computers to do my tests. My time allocation
>>code is the same if ponder is on or off.
>
>Then I'd bet the following happens: your average time per move is much larger
>at move 40 (assuming 40/2hr time controls) that it was at move 10. In my case,
>this is not true. I use more time up front, knowing that I will make up some
>of it with correct ponder moves that save time later on. IE I want to use
>more time from move 10-30, not get to move 40-50-60 and suddenly find myself
>with a huge wad of 'ponder time' that I saved...
>
>I wrote a paper for JICCA years ago (using time wisely) that explained how
>we did this in Cray Blitz. And we got the idea from watching lots of GM games
>to see how they used _their_ time. And they _all_ burn their time early when
>it matters, not in the endgame when it probably doesn't...
Well in this case I can maybe improve Tiger's time management a lot. My policy
is to try to get the same "apparent" thinking time all the time (time used for
each move on the real clock). This is probably why pondering on or off does not
matter much for me.
However my policy does not look so silly from what I see in the games. I often
see Fritz or other programs thinking much longer than mine in the begining of
the game, but they get nothing from it. Then they have much less time after the
middlegame, and are quite often _badly_ punished for this.
I also, a long time ago, removed from my program the "think twice as long on the
X first moves after opening book" heuristic.
Christophe
>>I would say that I almost never myself use my program with pondering on. My beta
>>testers use it, so at least I have a little feedback if suddenly something's
>>wrong with it. I have tested myself pondering=on just before the World
>>Championship, but it was the first time I was using it in several months.
>>
>>I don't care if people use my engine with ponder off in comp-comp games on a
>>single PC. I would even recommend to switch pondering off if you are using only
>>one PC.
>>
>>I would never take seriously a result where both engines were allowed to think
>>on opponent's time on a single PC. Doing so makes no sense, and I am always very
>>upset to see so-called "experts" postings results of comp-comp on a single PC
>>with pondering on.
>>
>>
>>
>> Christophe
>>
>>
>>
>>
>>>For _my_ program, "out-of-the-box" is the best way to run it, other than
>>>customizing hash table size for your specific hardware. Everything else is
>>>_exactly_ as I run it on ICC, which means that the 'defaults' are the best that
>>>I know how to do...
>>>
>>>Changing anything will very likely weaken it. Perhaps significantly...
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.