On Wed, May 13, 2009 at 2:34 AM, Christian Nentwich < christ...@modeltwozero.com> wrote:
> > Sparks are flying, but I don't think that either of you are onto the exact > truth. Don: do you play Go the same way you play chess? I don't think I do. > The "opening", in computer terms, for me only gets to move 20 if there is a > long Joseki involved. It happens too often that somebody "tries" something > before that. (Disclaimer: I only ever made it to 3 dan, and don't play > actively - strong players may think differently). Last time I talked to a > pro, many years ago now, about a relatively innocent opening (Chinese > opening, a pattern consisting of 5 moves), I found a very open mind about > each of the moves, and huge flexibility to deal with deviations. I'm only advocating an approach where you "remember" what you were thinking last. This is not inflexible, it is just the opposite. A strict opening book is inflexible. With this "memory" approach, you might find that what you think is good today, does not look so good tomorrow. If there is any lack of open-mindedness on the part of the program, the fault lies with the move selection algorithm, not the idea. In fact this is far more human-like than simply wiping your memory of each game after you play it. It's my feeling that on 19x19 this is not worth much of anything - certainly not the enormous disk space it would cost. It's not because of any inherent unsoundness of the idea, but just the fact that you are only going to be able to maintain a limited memory of just the first very few moves. There is no doubt you will be able able to play those very few moves much better, but it's not clear if some kind of joseki is better. > > I am pretty convinced that there is not currently a good answer to playing > well in the opening and early middle game and 19x19. I just don't think > anyone has come up with one. MCTS looks like a blind man wandering the > desert (I agree on this point), but on other hand a pure opening book will > limit the program. I am hoping that people keep a goal of beating the > *strongest* human players in mind, rather than chasing 200 ELO here and > there. It seems easy to tune programs that plateau at a particular kyu or > dan level. I think this much is clear. Even on small boards the program benefits from some kind of opening book. > > > In that respect I commend the authors of the papers on Monte Carlo go on > their reservations about hand-crafted patterns in playouts. They, the > patterns, do have a tendency to remind me of kyu player tactics (always > respond to atari, always pull a stone out of atari...). Pure statistical > measures (criticality, David Silver's approach), etc, sound *much* more > exciting to me, as they don't constrain strength scalability - I hope to see > many more. > > Christian > > > Don Dailey wrote: > >> >> >> On Tue, May 12, 2009 at 8:17 PM, Dave Dyer <dd...@real-me.net <mailto: >> dd...@real-me.net>> wrote: >> >> >> > >> >If I use persistent storage and do that search again in another >> game, I can start exactly where I left off and generate 50,000 >> more nodes. It will be the same as if I did 100,000 nodes >> instead of 50,000 nodes. Or put another way, it will be the >> same as if I spent 20 seconds on this move instead of 10 seconds. >> >> Sure, but except for openings you'll never reach the same >> position position in another game unless the players are >> deliberately copying their play. >> >> >> I think you still miss the point. I don't expect this to help on move >> 20, only for the early positions you have seen before. >> >> When I play chess, I play the first few moves almost instantly, without >> making a blunder. I can do this because I have studied those positions and >> know them thoroughly. It's not because I memorized them, even though I >> have. As soon as I get to positions I am less familiar with, I slow down >> to a crawl, because now I have to try to figure them out over the board. >> Of course my knowledge of the opening helps some, even when I'm out of the >> book I have a good sense of how to proceed. >> With GO, a computer can do the same. It's almost ridiculous to search >> the opening position from scratch, as it we are a totally moron and have >> never see that position before. There is no reason not to study it and >> remember what we learned about it. There is no reason not to apply this >> to as many moves as possible in the opening, even if it's only a very few. >> >> I don't understand why this is not immediately obvious to you. What I >> suggest is superior to just memorizing what moves to play, you can actually >> save the tree and in a sense it's like a program actually understands the >> position, not just play by rote. If it's a position you already searched, >> you will have at least a little head start no matter what the opponent >> plays. Even if the opponent surprises you, you will have a piece of the >> tree intact and the next time this same situation happens, you will now have >> some serious analysis already pre-computed. >> >> With 19x19 this is probably not a huge ELO booster, but there is certainly >> nothing unsound about it either. And it may not be as bad as you think >> because the computer will tend to be deterministic about which moves it >> plays which by itself drastically lowers the branching factor. If it fails >> to play the same in a position it has already seen, that is not a bad >> thing either, it indicates that it found a move it likes better, perhaps >> discovering that something was wrong with the previous choice. That's >> exactly what you hope it does. >> >> >> Consider move 20 (for example). If you saved every "move 20" node you >> ever encountered, how often do you think you'd encounter a duplicate >> from a different game, such that you can either avoid an evaluation, >> or improve your knowledge of that position by studying it >> a bit more. I contend it is a vanishingly small percentage. >> >> >> I don't believe you want to save the tree beyond the point where you are >> in a position you have never seen for the very reason you state. It's very >> unlikely that you will still be a part of the tree you have visited before. >> I thought I already conceded that point? Didn't I already say that >> this is an idea for the first few moves? >> >> But this idea will save you time that can be spend in later moves, so it >> can actaully benefit the moves you make later in the game. But more >> importantly it can prevent you from being in a losing position by move 20 >> from a bad move choice on move 5. >> - Don >> >> >> >> >> >> >> >> >> >> >> -- >> >> _______________________________________________ >> computer-go mailing list >> computer-go@computer-go.org <mailto:computer-go@computer-go.org> >> http://www.computer-go.org/mailman/listinfo/computer-go/ >> >> >> ------------------------------------------------------------------------ >> >> _______________________________________________ >> computer-go mailing list >> computer-go@computer-go.org >> http://www.computer-go.org/mailman/listinfo/computer-go/ >> > > > -- > > Christian Nentwich > > Director, Model Two Zero Ltd. > +44-(0)7747-061302 > http://www.modeltwozero.com > > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ >
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
