Re: computer-go: Neural Nets: suggesting and evaluating

[Erik van der Werf <E.vanderWerf@xxxxxxxxxxxxx>]

Peter Drake wrote:
> It appears that there are two obvious ways to use a neural network in a 
> very simple Go program:
>
> 1) Take the board as input and output a value for each potential move.  
> The move with the highest rating is the move considered best by the 
> network.
> 2) Take the board as input and output either (a) the value of the board 
> or (b) the value of each point on the board.  To choose a move, do 
> global search (usually 1-ply, conceivably deeper).
(1) I did something like that which is described in my CG2002 paper. One 
could also do this in a Q-learning framework.
2(a) In contrast this could be used in the TD-learning framework.

Both approaches seem valid.

2(b) isn't this the same as 1, or do you predict something else?


> Our current evaluator network has one input unit and one output unit for 
> each point on the board, and a number of hidden units.  Rather than do 
> deep thinking about how the hidden units should be divided into layers, 
> we let each hidden unit take input from all of the input units and all 
> previous hidden units, so that the hidden units form a sort of column 
> with weights coming from all previous layers.
>
> We had been looking at small windows on the board, but this has two 
> problems:
>
> 1) Vital connectivity and eye information may not be available within 
> the window.
Even if the full board is available your network will most likely not 
succeed in learning the long distance connectivity relations, at least 
if you use a raw representation. (Remember that Minsky & Papert showed 
connectedness to be of unbound order.)

My approach would be: Take a large window and extract a reduced number 
of features that emphasize the local structure.


> 2) A lot of time is wasted recomputing the same information over and 
> over again.
If it's really the same information why can't you just store it in a cache?

Regards,
Erik