Author: Vincent Diepeveen
Date: 05:57:46 05/14/98
Go up one level in this thread
On May 13, 1998 at 13:52:30, Don Dailey wrote: >On May 13, 1998 at 13:06:56, Christophe Theron wrote: > >>On May 13, 1998 at 07:14:16, Ralph Jörg Hellmig wrote: >> >>>So if there is a special time control, one program may play positionally >>>better, but the other one has better tactics, for example, the >>>positional better program will be stronger if the time control >>>increases, as it does also see the deciding tactics ... >> >>Another chess legend. >> >>Who has any proof of this statement? > >I have empirical evidence of it. That is if you mean longer time >(or faster hardware) favors the program with more knowledge. > > >>More knowledge better at longer time controls? Take a look at the top of >>the SSDF list, sit down a minute, and think again about this legend. >> >>I think that if you have more time to compute, you need LESS knowledge. >>We still have to find which kind of knowledge is needed in this case, >>and which other can be thrown out happily. >> >> >> >> Christophe > >Hi Guys, > > >I used to believe strongly that with faster and faster hardware, >knowledge becomes less important. The reason I believed this >was that eventually all programs would converge on a game theoretic >solution, which is essentially proof of this concept. > >HOWEVER, at the depths we are currently doing (and for the forseeable >future) it seems that the opposite is true. I did a big experiment >where many programs with varying amounts of knowledge played each >other. I generated hundredes of thousands of games on several computers >over several weeks of time. What happend was that the programs with >the most knowledge, improved very rapidly with depth compared to the >programs with little knowledge. > >I suspect with a great amount of depth, the knowledgable programs >would not be able to improve very much since they would be close >to "perfect while the dumb ones would be playing catch-up. But it >looks like we are a long way away from these ranges at current >time controls on modern hardware. > >About your reference to Fritz. Is Fritz really so bad at positional >chess? Some people confuse conservative play with bad chess. Could Fritz ain't that horrible, but it's in fact very horrible in seeing the position change. I claim that this is the result of piece square tables, but frans morsch says that he isn't relying on piece square tables. This could be true. This means it simply has no knowledge about what to exchange. This is also the big weak point of Rebel. Ed schroeder states here that he is relying on leaf evaluation. In that case, rebel doesn't have code about what pieces to exchange. Main point is that you may not evaluate the exchanges. You should look to the material that you keep afterwards. So the resulting position. I dunno why Rebel+Fritz+genius and some others are so horrible in these exchanges, but my first guess was piece square tables. Now i say: perhaps mobility is a reason too. The common thing of all these programs compared to Crafty, Zarkov, Diep and some others is that they lack mobility terms. It will remain an open question. Also pawn structure gets done wrong by Fritz, way more wrong than most other do it. Like in an endgame you have doubled a-pawn. Then fritz will try to capture these pawns at all positional costs. >this be the case here? It's hard for me to believe Fritz could be >that horrible and still be on top just due to a little extra speed. >I'll bet you will find that it's evaluation is reasonable, well >balanced and not as bad as it's reputation. It's my understanding >also that Franz has added knowledge gradually over time to keep up. > >The thing I notice about Fritz is that even on 1 ply, most of its >moves are reasonable, at least positionally. > > >- Don
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.