Magnus, Thanks for this detailed explanation. I guess AMAF plus Go knowledge based biasing is where the magic comes from. I feel a little comfortable now. I wonder where the sweet spot would be, ie, how many playouts would give us 80% of the effects. Quoting Darran:
>>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>> Incidentally I did some Fuego tests with just 300 playouts (9x9), and was surprised how strong it was in many positions (*). (This is basic Fuego 0.4.1, no extra pattern library, so relatively light playouts.) Darren *: My hunch is it would score 70-80% on the Realistic Test Suite. By "Realistic Test Suite" I mean a very large set of positions taken from real games (i.e. with no bias for difficult or interesting positions), with all winning moves marked and the program just has to choose any one of them. >>>>>>>>> >>>>>>>>> So, we are getting 80% of the effects for 100-300 of playouts? Regards, Fuming On Thu, Sep 9, 2010 at 6:37 PM, <[email protected]> wrote: > When the tree is expanded and we have no knowledge of go the statistics for > each move is initialized to 0 wins and 0 visits, as well as 0 AMAF visits > and 0 AMAF wins. > > With prior knowledge such as that we know that the last move can be > captured in a ladder this move can be initialized with for example 10 Wins > and 10 Visits. This makes sure this move will be searched until is has been > refuted or a better move found. The prior modification can be added to the > AMAF statistics, or to the "real" playout statistics for the move. > > Without priors the search will just play random moves until one seems good. > But for statistical reasons there will always be bad moves that get a high > win rate because of random variation and good moves get bad win rates. With > high quality priors the program will search moves that must be searched > locally at least making search sharper (without prior values for moves tend > to be the average of all moves searched before AMAF kick in and select > better moves) and robust against the randomness of the playouts so there is > less risk of missing to search the best move. > > When Valkyria expands the tree with a new node it will call its tactical > module for responses to the last move and add "prior wins and visits" to all > candidate moves of that node that has a tactical response value. Other > programs may call some slower pattern matching module to get high quality > priors for each move on the board. I think this is what makes many faces > strong. > > > Magnus > > Quoting Fuming Wang <[email protected]>: > > Erik, >> >> When you say "priors in the tree", do you mean the tree data inherited >> from >> calculations in the previous move? >> >> Regards, >> Fuming >> >> >> On Thu, Sep 9, 2010 at 5:43 PM, Erik van der Werf >> <[email protected]>wrote: >> >> On Thu, Sep 9, 2010 at 9:33 AM, <[email protected]> wrote: >>> >>> breaking even with gnugo on 9x9 at about 100 playouts), though I think >>> I get most of that from good priors in the tree. Maybe I should dust >>> off my old policy some time... >>> >>> Erik >>> _______________________________________________ >>> 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 >
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
