Consider the game when computer is black, with 7 stones against a very strong human opponent.
Computer thinks every move is a winning move; it plays randomly; a half-point win is as good as a 70-point win. Pro gains ground as computer makes "slack moves", taking slightly less than its full due. At some point, the computer, being the weaker player, makes one "slack move" too many and loses the game. Rinse and repeat. At some point, it dawns on the programmer: must attack to win handicap games. Must be a little bit greedy, to slow down the process of attrition. Dynamic komi models something real: the significant advantage of the computer in a handicap game. It tries to preserve as much of that advantage as possible. I don't know if it will work for computer vs human games. I do know that a similar idea helped me defeat a human player and reduce my handicap by 3 stones. Not having the patience for thousands of 30-minute games to achieve statistically valid results, I settled for trouncing my opponent three games in a row by a large margin, then doing it again with a smaller handicap for three more games. I can't win by 70 stones in a 7 stone game, but 20 or 30 was enough to prove my point. If I made random plays under the assumption that I still had a half-point win, my opponent's predictive powers would be superior to mine - that's why he gives me a handicap and not vice versa. I can't really be sure that my prediction of a 22.5 point win is exact to the last decimal point - but if it should be within 5 or 10 or even 20, I'm perfectly happy. It's nice that statistics of a series of one-bit values are so useful, but when a significant fraction of those one-bit values are 100% wrong, that introduces a bit of noise to one's estimates. One hopes that they balance evenly, but perhaps they do not. Terry McIntyre <[email protected]> “We hang the petty thieves and appoint the great ones to public office.” -- Aesop ________________________________ From: Don Dailey <[email protected]> To: computer-go <[email protected]> Sent: Wednesday, August 19, 2009 6:03:50 AM Subject: Re: [computer-go] Dynamic komi at high handicaps One must decide if the goal is to improve the program or to improve it's playing behavior when it's in a dead won or dead lost positions. It's my belief that you can probably cannot improve the playing strength soley with komi manipulation, but at a slight decrease in playing strength you can probably improve the behavior, as measured by a willingness to fight for space that is technically not relevant to the goal of winning the game. And only then if this is done carefully. However I believe there are better ways, such a pre-ordering the moves. Even if this can prove to be a gain, you are really working very hard to find something that only kicks in when the game is already decided - how to play when the game is already won or already lost. But only the case when the game is lost is this very interesting from the standpoint of making the program stronger. And even this case is not THAT interesting, because if you are losing, on average you are losing to stronger players. So you are working hard on an algorithm to beat stronger players when you are in a dead lost game? How much sense does that make? So the only realistic pay-off here is how to salvage lost games against players that are relatively close in strength since you can expect not to be in this situation very often agaist really weak players. So you are hoping to bamboozle players who are not not weaker than you - in situations where you have been bamboozled (since you are losing, you are the one being outplayed.) That is why I believe that at best you are looking at only a very minor improvement. If I were working on this problem I would be focused only on the playing style, not the playing strength. If you want more than the most minor playing strength improvement out of this algorithm, you have to start using it BEFORE the loss is clear, but then you are no longer playing for the win when you lower your goals, you are playing for the loss. - Don 2009/8/19 Stefan Kaitschick <[email protected]> One last rumination on dynamic komi: > >The main objection against introducing dynamic komi >is that it ignores the true goal >of winning by half a point. The power of the >win/loss step function as scoring function underscores >the validity of this >critique. And yet, the current behaviour of mc bots, when either leading or >trailing by a large margin, resembles random play. >The simple reason for this >is that either every move is a win or every move is a loss. >So from the >perspective of securing a win, every move is as good as any other >move. >Humans know how to handle these situations. They try to catch up from >behind, or try to play safely while defending enough of a winning margin. >For >a bot, there are some numerical clues when it is missbehaving. >When the >calculated win rate is either very high or low and many move candidates have >almost identical win rates, the bot is in coin toss country. >A simple rule >would be this: define a minimum value that has to separate the win rate of the >2 >best move candidates. >Do a "normal" search without komi. >If the minimum >difference is not reached, do a new a new search with some komi, but only >allow >the moves within the minimum value range from the best candidate. >Repeat this >with progressively higher komi until the two best candidates are sufficiently >separated.(Or until the win rate is in a defined middle region) >There can be >some traps here, a group of moves can all accomplish the same critical >goal. >But I'm sure this can be handled. The main idea is to look for a less >ambitious gloal when the true goal cannot be reached. >(Or a more ambitious >goal when it is allready satisfied). By only allowing moves that are in a >statistical tie in the 0 komi search, >it can be assured that short term gain >doesn't compromise the long term goal. > Stefan >_______________________________________________ >>computer-go mailing list >[email protected] >http://www.computer-go.org/mailman/listinfo/computer-go/ >
_______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
