I think you are essentially correct and understand the issue here. I think of it in terms of risk vs reward. The bot is all about trying to insure a victory, not providing a safety buffer. its safety buffer isn't more points, but making just the required number of points needed safer. Strengthening an already won area even more is always less risk than getting some points on the board it doesn't need.
Don On Thu, Jun 6, 2013 at 11:44 PM, Mark Goldfain <[email protected]>wrote: > Forgive me if this is too basic an observation that most all of you > already know full well. > > It seemed there was a question out there as to why MCTS bots often win by > a very narrow margin. And I did not see an answer such as this posted > (again, perhaps because everybody knows this). But if I am correct, then > at least I'd like to know that. I have not programmed such a bot, so I > would like it if one of you affirmed or refuted this. > > I believe it is a simple interaction between the goal for these bots and > the nature of the game of go. The bots, as has been pointed out, care > nothing for the margin of victory. They will count a 0.5-pt win as just as > good as a 200-pt win. Put that together with what we all know as human > players: There are always moves out there that eliminate a cut point, or > secure a given territory beyond all dispute. We spend the whole game > knowing about these moves, but not making them, because we have a higher > priority. Our well-matched opponent is doing everything he/she can to push > us and outrace us in achieving the building of live shapes containing the > maximum number of points. > > If we were really good at reading out the rest of the game and were able > to confidently say we were 10 points ahead, and we did not need to make the > maximum point-gain at every move, we might be able to make one or more > "insurance" moves. I think the bots are led naturally by their > calculations to make such moves. > > This is not how human players play. I think human players learn to mostly > always consider what next move will be of the most value. They do not > fluidly switch when they are ahead to conservative moves, because we are > never certain we are ahead until it is pretty clear. I was recently being > thrashed by a much stronger player. I think he was up by about 40 points. > What did he do next? He knew full well that his groups were safe enough. > He attacked and reduced one of my territories, and went on to have about a > 55-point margin of victory. I think that a bot, looking at the > probabilities of victory, would instead have been led to take it easy on > me, and go for a result that was closer, but left no chance for any attack > on my part. Oddly enough, I know some human players get very angry about > this behavior, they do not generally consider it as a polite way to act. > I've heard comments like "The bot will just toy with you, giving you a > feeling you might yet win." If I'm correct above, it is nothing of the > sort. > > ______________________________**_________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/**mailman/listinfo/computer-go<http://dvandva.org/cgi-bin/mailman/listinfo/computer-go> >
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
