On Wed, 2007-02-28 at 10:12 +0100, Heikki Levanto wrote: > On Wed, Feb 28, 2007 at 09:28:37AM +0100, Magnus Persson wrote: > > >I have never really understood the idea in all-in-first. > > > > Actually early in the game the order of random moves does not matter. The > > Sequence AbCdE... have same result if for example A and C switch order for > > CbAdE... If the sequence is a win then standard MC will attribute this to > > playing A first, but actually the sequence with move C first would also have > > been a victory. hence all moves in the sequence could in theory be > > reinforced. > > I buy that for the first moves. But every game ends in filling points at > the edge, and giving them equal value to the opening moves sounds > counter-intuitive to me.
What I do in AnchorMan is that I don't go all the way. I run a simulation and only go about half way (I think it's 5/8 but I haven't looked at the code in quite a while.) I take 5/8 of the length of each simulation so some simulations may count more moves than others. I also only consider a move if the computers side was the FIRST to play that move, which deals somewhat with captures changing things. One can imagine other conditions too. Another possibility I tried but rejected is to consider each 3x3 pattern unique. So for an empty point, I keep stats only if the surrounding pattern matches the initial position being searched. This seemed like a great idea because it dealt with interference of move ordering fairly well, but it didn't test very well and it slowed things down a lot. - Don > Perhaps, if the weight given to a move would decrease in proportion to > how far ahead the move is, or something similar. > > Actually, I think my ownership map could be used to do a comparable job > - it also records where the stones ended in the game, which is a good > approximation of what moves have been played. Not quite sure how to use > it, yet. > > -H > _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
