Quoting Jacques Basaldúa <[EMAIL PROTECTED]>:
*Question* : 1.How does MC/UCT find the correct answer if the more you approach it (except if you hit it by a fluke), the worse it evaluates?
It does not find the correct answer of long ladders in general. I think the common behavior of a pure uniformly random MC-program is to play out ladders no matter if they work or not. The reason is probably that random moves might intervene and switch the status of the ladder several time before it is actally played out. This means that atari will always increase the probability of capturing the stones a lot in the random game.
2. Is UCT biasing the distribution of moves to be searched *against* the only correct answer i.o. towards it?
Plain UCT does nothing about this unless it searches deep enough. Adding tactics knowledge to the simulations makes a difference however. I found that even knowledge that it is far from reading out complete ladders, actually can make the program behave as if it was able to do that. It might be that the selective search of UCT can go really deep if it just gets a little help. If I understood the Mogo paper correctly their only tactics is to save stones that are in atari and can be saved and that seems towork very well on 9x9 although my feeling is that when Valkyria beats Mogo it is often because of tactical mistakes by Mogo. -Magnus _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
