I think we should open up to other ideas, not just dynamic komi modification. In fact that has not proved to be a very fruitful technique and I don't understand the fascination with it.
First we identify what it is we are trying to accomplish. You mentioned improving the strength of MCTS go programs in handicap games, but many people point to the ugly style of play. Here is something to think about in more theoretical terms. Define what good play is. To me there is no satisfying definition - the CORRECT definition is not the one that we intuitively expect. For instance, if you are in a dead lost position, every move you play is a losing move. How can you say any particular move is better? To me a good move is any move that preserve the best game theoretic result possible at this point in the game. That means in a lost position ALL moves are good moves because they all preserve the loss. It's an odd way to look at it, but it makes more sense in won positions, as only a few or even 1 move preserves a win. The intuitive definition that we really mean when we talk about good play is to play in such a way as to increase our chances of winning against fallible opponents. In fact, what is good is technically (and practically) defined by WHO your opponent is. It appears that dynamic komi modification does not even help you win against fallible opponents, so we should probably look to opponent modeling (since that is really what we are asking for - especially with handicap games.) Another definition for good play is one that could be added to our MCTS programs. Play the move which preserves the best possible total point score. That means even in a dead lost position we will fight for every point on the board. Why doesn't this code up well with go programs? From a purely theoretical view, it is a "perfect" goal in the sense that if it is followed perfectly, it will perform the same as simply playing for the win against perfect opponents. Part of the answer may be that it requires more skill. There are fewer moves (in general) that will accomplish this goal than there is in winning the game. It would not suprise me, although I haven't specifically tested this, if this led to stronger play than dynamic komi modification. Then there is the possibility of a hybrid between the two. A fourth possibility is to gently impose some higher order knowledge on the search - try to make the tree branch out with moves that WE consider stronger (by the practical defintion, not the correct definition.) Probably with only a little help, moves which are other equal but we like better can be coerced into being the ones expanded on first and thus end up being the ones chosen by the MC program. We could do this by letting some conventional program choose the best move(s) and give them a few bogus wins to make the tree begin with the moves we would prefer. There a probably tons of possibilities. What does Many Faces do? Does it play more naturual? I ask because I know that Dave is more concerned about the marketability of the program and is interested and concerned about such things. - Don On Sat, Jul 11, 2009 at 10:26 AM, "Ingo Althöfer" <3-hirn-ver...@gmx.de>wrote: > One of the difficult questions is if (or better how) > dynamic komi can be used to improve the strength of > MCTS go programs in handicap games (both cases being > "interesting": computer on strong side - and - > computer on weak side). > > Especially, there are several normal go players > ("non-programmers") who are interested in this topic > and who feel that the current non-activity on this > question is unsatisfactorily. > Stefan Kaitschick (German 5-dan amateur, EGF-rating > 2439) is in some sense the most prominent amongst them. > > I think that it would be a real chance for the programmers > to use the interest and creativity of such people. This might > happen for instance in the following way: include a > > "simple komi-modification button" > > in your program (or a pair of buttons, one for +, one for -) > which can be used by simple mouse clicks. Of course, all the > time the true value of komi should also be shown. You programmers > would probably be surprised to see what creative users would find > out when playing around with such a feature. > > Ingo. > -- > Neu: GMX Doppel-FLAT mit Internet-Flatrate + Telefon-Flatrate > für nur 19,99 Euro/mtl.!* http://portal.gmx.net/de/go/dsl02 > _______________________________________________ > 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/