[computer-go] (no subject)

["Pierce T. Wetter III" <pierce@xxxxxxxxxxxxxxxxxx>]

On Jun 8, 2005, at 4:04 AM, Thore Graepel wrote:



> Hi,
>
> I would like to let you know about a paper on move prediction using
> Bayesian pattern ranking that we just submitted to the NIPS  
> conference.
>
> http://research.microsoft.com/~thoreg/papers/gopat-draft.pdf
>
> Our system pretty much uses the same local move patterns as Frank de
> Groot's Moyogo studio (without liberty information) but applies a
> sophisticated Bayesian learning algorithm to estimate the values of  
> the
> patterns. Training was performed on 21000 games from the GoGoD  
> database
> and we provide move prediction statistics on 500 test games. Our  
> results
> appear to be slightly better that Erik van der Werf's results  
> (tested on
> 50 games and based on a different, more extensive feature set  
> including
> "location of previous move") but (not surprisingly) fall short of  
> Frank
> de Groot's preliminary results (tested on a single test game and based
> on about 20 times more training data as well as more sophisticated,
> context-aware small patterns).

 Hmm... That's very interesting. Would you be all interested in  
applying
that method to GnuGo? It has a couple of places where machine  
learning like
this could be very well applied.

 Specifically, it has some move ordering things that already use  
weights.

 Pierce


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