This type of behavior of course is unnatural for human players and even offends our eyes so to speak, but it's perfectly logical.
Keep in mind that for a program to do this, there has to be NO difference between the winning odds of playing G1 or K1. When a program does this it's not being recklessly foolish. It's actually being conservative and pragmatic. In your example, I can image that if you could make it favor the greater win when it clearly makes no difference whatsoever, you might increase the playing strength in a very minor way. You would be making is slightly more robust in the face of miscalculations. I think the right approach was mentioned just recently, add a very tiny incentive that cannot effect any bits of the main calculation as a kind of tie-breaker. I doubt this will buy much, but it might help just slightly and make it appear to play more "natural" by our standards. But I know trying to fix this any other way reduces the playing strength significantly. - Don Lavergne Thomas wrote: > On Tue, Oct 30, 2007 at 08:07:49PM +0100, Heikki Levanto wrote: > >> On Tue, Oct 30, 2007 at 02:45:56PM -0400, Jason House wrote: >> >>> Similarly, I've been in won games and gotten bitten by a tesuji by the >>> opponent. If I had been just a bit safer in my play, I could have had a >>> comfortable win. Similarly, reasonable MC bots solidify the core to win >>> rather than try to keep everything. >>> >> I would like to think so too, but that is not what I am seeing. In my own >> small experiments, I saw this kind of things often enough (bottom edge of the >> board): >> >> 3 O O O O O O O + + + + + + >> 2 O X X X X X O O O O O O O >> 1 O X + X + X + X X + X X O >> A B C D E F G H J K L M N >> >> Any "sane" human player would connect at G1 first, and K1 later. But to a MC >> player, those two are very close to equal, as long as the game seems to be >> decided by five points or more, or there is another similar situation on the >> board, and the program can be sure of getting one of them. >> >> > > This is expected for a monte-carlo evaluator. Let assume that any other > play on the board cannot change the result, and black is more that 5 > points in advance, you can play either move theresult will be same : a > victory. The only thing that will change is the score of the game, but > in most MC engine the score is not used, only the result. So you can > make all simulation you want, the score for g1 and k1 will be same : 1.0 > > MC tend to choose the safer path in the tree. At a given node, if you > have only 2 possible moves : > - the first one is a clever tesuji which lead to a 50pt win but need > to folow a strict sequence of 10 moves or you loose, > - the second one assure you a small victory of 0.5pt even if you make > some mistakes later in the game. > Your program will choose the second move because more random simulation > lead to a win in the second case. > > The first type of move is the ones that you can look in professional > games where pro are always on the line and any mistakes can make you > loose the game. Second one reflect the safest way of playing when you > have some advance in the game, and when you just want to win the game. > You just make your group fully alive and solid, and you prevent any > invasion from your oponent. > > Tom > > _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
