[Computer-go] A cautionary tale about randomness, Part 2

Dave Dyer ddyer at real-me.net
Fri Sep 7 19:45:29 PDT 2012


>
>Are there performance reasons that made you try anything different?

That's the whole point.   Ideally, if the position has 100 legal moves,
generating 1 ought to be 100x faster than generating them all and 
throwing 99 of them away.

Depending on the absolute cost of generating a move, this can be
very significant.  

My implementation of Hex for example, the number of simulations
only increased by about 2x, but the randomization was trivial
and perfect.

For Blokus the simulation rate increased by 10x, and the program
got a lot stronger, even though I now know that my randomization
is really bad.





More information about the Computer-go mailing list