On Wed, Jan 14, 2015 at 1:32 PM, Petr Baudis <[email protected]> wrote: > > Note that the mistake I did with Pachi is that I worked hard to optimize > the Go board (playout) code with minimal functionality to perform > a playout correct by the rules, only then discovering how much other > information I need to keep to add some useful heuristics to the playout, > and maybe the hotspots change then. > > If I did it again, after picking some really low-hanging optimization > fruit I'd rther focus on implementing playout heuristics, only then > coming back to optimize the board code when it is clear what all > information I need to compute the heuristics efficiently (list of up to > N libs, maybe pre-computed pattern hashes, etc.). >
That's a good point Petr! But as I started with 50pps on 9x9 (yes, really) I actually needed to do some optimisations as this made the bot completely unusable. There are still a few low hanging fruit (or learning opportunities for me), but it's probably time to finally start implementing UCT. BTW, the board.c from pachi is very readable and it (together with Board.java from the current Orego) serves me well as an inspiration on how to design the board. Urban
_______________________________________________ Computer-go mailing list [email protected] http://computer-go.org/mailman/listinfo/computer-go
