Matt Gokey wrote:
But what are some of the reasons MC is not even better?
1-Since MC engines don't deal with tactics directly, they're not likely
going to play tactical sequences well for low liberty strings, securing
eye space, cutting and connecting, ko fights, or ladders, etc.
2-Also because most of the play-outs are usually nonsense, they may
have trouble dealing with meaningful nuances because the positions that
will lead to these distinctions just don't arise with enough statistical
frequency in the play-outs to affect the result. Yet when very
selective moves are used in the play-outs, too many possibilities can be
missed.
3-Finally, with 19x19 anyway, the size of the board and game tree
probably limits the practical effectiveness of the sampling and move
ordering. I don't try to address this last point any further in this
message.
Very good analysis and I would like to contribute a 4th reason:
As Luke Gustafson pointed out, we have to contemplate the simulation
as a _stochastic process_. We want to determine the conditional
probability of a win given a particular move is made. And that depends
on the _length of the simulation_. Dramatically! This is a reason
against scalability of global search MC/UCT. If the simulation is
500 moves long (Chinese rules with recaptures, etc.) the observed
variance at an early move blurs out everything.
Just a simple stochastic process: Count a dollar each time you
correctly predict a p=1/2 coin thrown n=500 times. The expected
average is (obviously) 250, but the expected variance of that
measure is n·p·(1-p) = 125 proportional to n.
Jacques.
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
