On Mon, Oct 26, 2015 at 5:03 PM, Gonçalo Mendes Ferreira <[email protected]> wrote:
> I think we mean the same. So when black plays at a1 first we count it for >> black everywhere but never for white. Correct? >> > The question is, if both black and white play a1 - because there were > captures in the middle - whether both of them will update a1 in some of > their nodes. I believe so. I see. Yes, that is also an option. I’ve only checked Michi and it only counts the first one. It’s possible that this is just out of performance considerations, but it could also be argued that so many captures happen in the playouts that the information at the end of the game (where captures are more likely to happen) isn’t worth recording. > > Ah yes, the UCB term. I’ve now added it. I dimly remember now that Pachi >> doesn’t use it but yes I’ve now added it. And I’m not sure I understand >> you >> other comment. My nodes store the visits/etc, the move made, and >> references >> to all children. >> > I've checked and that's with Pachi/Michi do too. And it's the closest to > the MCTS diagrams and explanations that are published around. But instead > you could just have the information on transitions from a state, and not > wins/visits of the state per se. I'm not very familiar with Pachi (I > couldn't find how it implements a transpositions table, or how it prunes it > with memory safety). > > I use a structure with > color > visits[19][19] > wins[19][19] > amaf_visits[19][19] > amaf_wins[19][19] > etc > > They are very similar, but my node information doesn't need to consult the > transpositions table or follow points to select the next play. On the other > hand sequences of play that arrive on the same node have the exact same > "quality" regardless of sequence of play in Pachi, where in my program the > sequences of play only "unite" when following that state forward. > > I don't know if this difference is very small or if I'm forgetting > something important. Maybe Pachi implements discovering if a state has > already been expanded in a very efficient way. > > I don't know how other programs do it, but instinctively, if you started > by implementing a non-UCT MCTS algorithm, the information you kept would be > just the initial state and statistics on the transitions, so my train of > thought seems (to me) the logical follow up. > I must admit that I don’t quite follow how you’ve implemented things, but then you seem to be further along than me as I haven’t even started with transposition tables. Urban -- Blog: http://bettong.net/ Twitter: https://twitter.com/ujh Homepage: http://www.urbanhafner.com/
_______________________________________________ Computer-go mailing list [email protected] http://computer-go.org/mailman/listinfo/computer-go
