On Mon, Jun 3, 2013 at 9:14 AM, Stefan Kaitschick < [email protected]> wrote:
> > > > My original question was meant for non-resigning bots. The observation: > > Also a losing bot starts to play lazy towards the end. And the question > > is: > > > > To which results do laziness of the winning bot AND laziness > > of the losing bot together lead typically? > > > > Ingo. > > > A losing bot is never lazy. It can get despondent though. :-) > The most interesting result is that a slight bonus on higher margins > can improve the success rate, even though it might slightly reduce the > bots internal winrate. I would venture the hypothesis, that increasing > the margin from 0.5 to 1.5 is worth the greatest winrate "investment" > by far. Messing up the final ko is always a possibility. > This basically comes down to opponent modeling. You can improve the success of a bot by trying various things to make it play more aggressively, but in some theoretical sense you are reducing the strength of the bot. In a similar way I can beat weaker players at chess more easily by playing unsound moves. It is possibly the case that you could even improve the play in a real sense if such tricks address an actual weakness in the program, such as providing a buffer against poor understanding of certain endgame shapes (Nakade.) That is like a "band-aid" fix but not necessarily a bad thing. We do whatever works. There are basically 3 reasons to use such tricks: 1. To improve the style of a program for cosmetic reasons. 2. To improve the results against other players (opponent modeling.) 3. To get an actual improvement in strength with a general band-aid fix (as opposed to addressing the actual issue.) In handicap games one side basically starts from a dead lost position, so even the top players rely heavily on opponent modeling. For a strong bot you would want the program to "stay in the game" until the opponent makes enough mistakes to make a win possible. Don > > Stefan > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
