Re: [computer-go] Learning : was Chess programs versus go programs

[Nic Schraudolph <compgo@xxxxxxxxxxxxxxx>]

> The more complexity something has, the more difficult automatic tuning 
> it
> gets as the optimizers don't even know where to start optimizing.
True to some extent, but humans get lost *far* earlier than learning 
algorithms.  There are psych studies showing that the average human 
cannot handle more than 3 nonlinearly interacting parameters 
simultaneously, and trained experts peter out at 5-6.  You can handle 
more only if their interaction is semi-linear so that the tuning 
problem decouples nicely.

Now, some optimization methods do not perform well past a few hundred 
parameters, but others can handle millions. There are many unsolved 
problems in learning computer go, but the optimization is not one of 
them.

- nic

--
   Dr. Nicol Schraudolph             http://n.schraudolph.org/
   National ICT Australia

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/