Author: Michael Yee
Date: 09:21:06 05/10/05
Go up one level in this thread
On May 10, 2005 at 10:17:39, Jay Scott wrote: >On May 09, 2005 at 22:49:37, Michael Yee wrote: > >>I came across this interesting/strange paper today: > >The results don't fully support the claims, which is bad but unfortunately >common. And, like other work by David B. Fogel, the paper doesn't show much >knowledge of past work and of the state of the art in game playing and game >learning. But the paper shows strength in the use of machine learning--chess >programmers who are trying these techniques out make a lot of mistakes which are >avoided here. > >I thought the paper was interesting. We can learn something from it. But don't >believe everything it says. For me, the main lesson was that parameters can be tuned through self-play and that beneficial "positional" features can potentially be constructed/represented by neural nets. >>(1) started with known good values for material and PSTs (so did in fact >>incorporate human knowledge) > >Avoiding use of human knowledge does not seem to have been a goal. Given that, >starting with best-guess values is an effective way of speeding up learning. While I agree that using good starting points makes things more efficient, to me they seem to play up that they use no expert knowledge (at least at the digenetics page regarding their "chess with an attitude" program and also in Fogel's numerous checkers papers--where I feel they had features relating to material already given to the network as input). In the chess paper, they mention a goal of learning without "significant human intervention". But in my opinion they applied a tremendous amount of domain knowledge in this case. We don't even really know how sophisticated the base engine was (e.g., maybe it had an amazingly efficient search that required only a reasonable evaluation function to work well). >>(3) had artificial rule that games lasting 50 moves were draws (although I can't >>tell if this was just for during training) >>(4) fixed search depth of 4 ply (or 6 ply when extending for quiescence), but >>maybe only during training? > >The learning algorithm is extremely slow. Long searches and long games would >have wasted a lot of cpu time. > >>(6) cites Kasparov communication as a reference... > >That was depressing. :-( Name-dropping for an obvious suggestion. > > Jay Thanks for the comments. The paper was a bit frustrating and I'm glad it wasn't just me :) Michael
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.