Author: Chris Whittington
Date: 08:56:39 11/13/97
Go up one level in this thread
On November 13, 1997 at 10:42:40, Randolph S. Baker wrote: >On November 13, 1997 at 09:16:27, Graham Laight wrote: > >> >>On November 13, 1997 at 07:13:03, Chris Whittington wrote: >> >>>This is not directed specifically at Amir ....... >>> >>>I think the whole lot of you are avoiding the crucial issue from the >>>games at WMCCC. >>> >>>The fast searchers, even with 767 alphas, were expected to sweep the >>>board. Manifestly they didn't. >>> >>>Some other fast searchers, running on PC's also under-performed >>>according to expectations. >>> >>>Several programs (ranging from very slow, to quite fast, but none of >>>them brute monsters) were not even spoken about before the WMCCC as >>>being of any interest, performed way above expectations. >>> >>>One program (self-promotion prize Kim-il-Sung already awarded) running >>>at 4000 nps did really rather well. >>> >>>Something is going on, and none of you is addressing it. >>> >>>Compare the cock-crowing and hubris from before the event .... >>> >>>The old knowledge-speed issue which gets jumped on as >>>boring/tedious/been through it all before/our way is best by the usual >>>culprits rears its ugly head again :) >>> >>>Chris Whittington >> >>If increasing computer hardware speed is tipping the knowledge/speed >>battle in favour of knowledge (which seems to be the prevailing >>doctrine), then one could expect the following effects: >> >>If a clever program played a quick one at low time controls (e.g. game >>in 5), the quick ones should win. >> >>At medium time controls (e.g. tournament chess times), it should be >>getting more even. >> >>At long time controls, the clever programs should be dominant. >> >I observed a very similar pattern in playing Fritz5 versus the Hiarcs6 >engine, with 1 additional discrepancy which also makes sense: at very >short (relatively) time controls, search wins over knowledge because the >knowledge programs miss tactics. (My testing was done on a P90, so some >appropriate scale applied to the time controls compared with a P233 >would also account for it). > >In very fast games (e.g. 4'+2'' or 5' fixed), Fritz5 wins handily. At >modest blitz speeds (5'+12'') Hiarcs6 was even or slightly better. At >somewhat longer time controls (25' fixed), Fritz5 was slightly better. I >haven't run many tournament time matches, but I have 20 game Nunn match >at 40/2 in progress. Results for first 6 games are Fritz5 +3-2=1, (too >close to call.) > Unfortunately, testing with hiarcs and fritz like this does prove much. the alternative explanation for the results swing is that at longer time controls fritz eats hash table, runs out of ram and stalls. Hiarcs only experiences this phenomenon later due to its lower node rate. To account for hash effects, you'ld need to run the experiment with several different hash sizes, including soem where both prgs had seriously large amounts of hash ram. You should then see the fritz results improve with more hash ram at long time controls. You may also then be able to pick out of the mass of figures soem data to show hiarcs getter better with longer time controls and 'normalised' hash figures ..... Chris Whittington > >>So - if Tal was to play Fritz at a rate of 1 hour per move, would it >>have a better chance of winning? It should have. According to Chris, Tal >>should be the perfect program for correspondence chess. >> >>However, I have my doubts and suspicions as to whether it really would >>dominate at long time controls. I agree with the theory, but I have >>nagging doubts as to whether practice would agree with the theory. Has >>anyone done any testing along these lines?
This page took 0.01 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.