Author: Gian-Carlo Pascutto
Date: 02:13:44 09/17/00
Go up one level in this thread
On September 16, 2000 at 20:05:01, Dann Corbit wrote: >On September 16, 2000 at 10:51:34, Gian-Carlo Pascutto wrote: >Depends on how you define it, I suppose. Alpha-Beta isn't really pruning, >since it always gives the identical solution, at some point. It cuts of parts of the tree. That's pruning isn't it? (I think you have forward pruning in mind: cutting of parts of the tree when you cannot be certain that cutting of that part is safe) >>As a more practical example, Hiarcs ply depth and Fritz ply depth >>are comparable? > >Yes. They clearly are comparable. A linear function would probably provide an >exact equation to fit them within a very close tolerance. Aha! So solving the challenge with Hiarcs to 16ply would be quite different to solving it with Fritz to 16ply. >Whether you know it or not is unimportant, I suppose. To search to a given >depth, through whatever means, is necessary to find a good move. All programs >search in a very similar way and have the same exponential algorithm behavior. >That's because the problem is an exponential problem. At some point in the >search it will have considered all the important moves out to a certain depth >in plies. The algorithm may not be able to determine what that depth is, but >that search will have happened. This can plainly be seen from first >principles. I agree with you here. But I don't think this jusitifies comparisations between those depth's. >Obviously, extensions greatly improve a program's strength (if done properly). >I think Phalanx is a model that many people might want to consider in this >regard. It consitently finds a correct solution at an earlier ply because of >the extensions employed. If it 'consistently' finds correct solutions at earlier plies, doens't this more or less defeat your whole argument? >>Also...remember Steve Ham's games. What did he note about the programs >>evaluation? I'm not saying that is _proof_ that you are wrong, but it's >>a good indication. > >I don't think the way humans play chess and computers play chess are very >comparable. I could be wrong, of course, but first we'll have to have someone >discover how humans do it. Nobody knows yet. If Steve says: this is a draw, and the program says +1.00 for me, and ten moves later they agree a draw, then Steve was obviously more correct than the program, even though the latter played the correct moves to arrive to the draw. (Or maybe it blundered away it's win and Steve was right...ok I admit Steve's judgement was a weak arument) >Chest gives a certificate of proof, and will even expand it to prove the >solution if you ask. Someone did offer a chest certificate for the position in >question, and I have accepted that solution. Full solution tree...I'll remeber that ;) >>I never said you did. I just demonstrated the approved measure of the >>PGN standard is shaky. > >And yet you use it. You berate me for my lack of mathematical rigor, and yet >you use each and every measure you chide me for in your program. Why doesn't >your program just return a -1, 0, or +1? The new search does. (It has an UNKNOWN result too though) >That's the only certain mathematical result. But it is totally impractical >except for rare situations (when a mate has been found). Agreed. >I agree with that, of course. But how do you decide about "the correct move is >played?" You certainly don't find the mate/loss/draw 99.9% of the time in a >search (except at the end of the game). You (rather) use a statistical >estimate of material and position to decide a value. Whatever you decide to >use, if your search is consistently successful, there will be an approximate >mapping from your search to some other successful program's search. Is it pure >chance that the WAC epd test suite will get approximately the same answers from >top programs at similar time controls? Most of the positions are not >checkmates, so they are finding "the same solutions" without any pure >mathematical rigor of true >solutions. But somehow they get the correct solution anyway. They might get the same answer, they won't get the same eval. And we can't be sure about the answers either. I remeber when Bob once complained that his program was scoring less at some test suites because the TB's were improved. He was seeing provably correct moves (mates), whereas the reported solution was one most other programs had agreed on. >>Louis Victor Allis, Searching for solutions in games and artificial >>intelligence, PhD thesis > >Is this used in Chess? Definetely. >>>I don't think so. Each ply (in my definition) is a half-move carefully >>>examined by the program [in such a way that the optimal chances of winning >>>from performing the analysis in that manner are performed]. >> >>Now it gets even better. I can show you Sjeng is able to search to >>depths of 100 000 000+ plies. I even let it go to 300 000 000 once. >> >>This definition is equal to what most people call a 'node'. > >Then your program wasn't playing chess. There are less than 300 million moves >and responses allowed in a game (see Steve Pribut's Chess FAQ). My program examined over 300M half-moves carefully so the chances of winning due to making the analysis in that matter were maximized. Fits your definition. >Maybe that is one more reason computer chess programs are hard to understand. >Everyone uses any term they like to mean anything they want. Unforunately true, but I used your definition. >A ply is the legal moves possible from a node. The next ply is the set of >responses to the choices made. This is a different one that the one you posted rearlier. Now I don't do 300M plies anymore ;) -- GCP
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.