computer-go: Neural Nets: suggesting and evaluating

[Peter Drake <drake@xxxxxxxxxx>]

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).

We're currently doing 2(b), but it's clear that this will not scale up 
to larger boards or deeper searches.  I'm contemplating combining these 
two approaches, so that one network suggests a move and another network 
evaluates the resulting board.  The suggestor could then be used to 
narrow the search in a number of ways, including:

- Consider the N highest-rated moves, such as the top five.
- Consider all moves whose estimated value exceeds some threshold.  
This threshold might increase with search depths, so only obviously 
good moves would be considered in deep searches.  Such a system might 
conceivably be able to read ladders as a special case.

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.

2) A lot of time is wasted recomputing the same information over and 
over again.

In light of that second point, evaluator and suggestor networks should 
probably share the same hidden units...

One of the oft-mentioned advantages of neural nets is the speedup due 
to massive parallelism.  We never actually get this, because we 
simulate the networks on serial machines.  Has anyone taken advantage 
of massively parallel hardware for speedup?  I believe there are 
generic neural net machines available.

Thoughts?


Peter Drake
Assistant Professor of Computer Science
Lewis & Clark College
http://www.lclark.edu/~drake/