On Mon, Apr 1, 2013 at 1:10 PM, Jason House <[email protected]>wrote:
> On Apr 1, 2013, at 11:10 AM, "Gabriel .Santos" <[email protected]> > wrote: > > > Hi! > > > > I'm a new computer-go researcher and I'm not a Go player. In order to > get better knowledge of Go Game I would like to ask some questions about it > (I know the rules of the game, I'm just not a good player). > > > > 1 - in order to evaluate simulations in MC. Is there any connections > between the type of moves made in the game ? For example, if i take two > simulations, victory in both simulations, in one of them I had just one > nakade move and in the other one I had 5 nakade moves. Could I say that > simulation two is better than simulation one ? By better i mean is it more > worth that I take more time simulating the states from the second > simulation instead of the first one ? > > I do not know of any engines that differentiate between the quality of > simulations, only the result. The investment in a particular tree node is > based on the win rate, the rave win rate, and bias with priors. > Of course that's no reason not to try it but it seems like it would be a really difficult proposition. If I understand this I think the point is that perhaps there is more relevant information contained in one playout over another and somehow it might be possible to take advantage of that? Don > > > > 2 - So, in this way could I conclude that, for example, Nakade moves are > ALWAYS better than Atari Defense Moves ? > > I think there are very few black and white rules about which heuristic is > better than another. There are a few different approaches to use heuristics > inside a playout. Most are statistical. > > > > 3 - As far as I know the alpha-beta approach has not succeeded due to > the inefficiency of the evaluation functions known. So,where do you guys > think that lies the future of Computer-GO ? MC methods ? The classic > approach on board games ? (Minimax, Neural Networks, etc). > > MC is definitely the future. I think there are ways to blend classic > methods with MC methods, but most are still experimental. > _______________________________________________ > 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
