Re: [gnugo-devel] Re: gnugo-devel Digest, Vol 13, Issue 5

[Gunnar Farneback]

David wrote:
> GNU Go is quite large and we keep trying to treat it as a "black box"
> move generator and evaluator. It keeps sucking us in deeper.

I would have tried to do everything through the GTP interface. That
should give a good isolation against the internals and allow easy
upgrade to newer versions. Possibly you might need some extra GTP
command but unless it's too exotic it would probably be accepted into
the main line.

> 3 of the lookahead paths are the top 3 moves selected by GNU Go, the 
> 4th is the top move selected by GNU Go if we were to pass and Black 
> were to choose (unless this is an illegal move for white, in which case 
> the 4th choice is the 4th highest move for white). We often choose to 
> take "their" best move over our best move. In other words, we are 
> testing the path suggested by the Go proverb "Your opponent's move is 
> often your best move."

My experience is that this proverb isn't correct as often as one would
like to think. It's natural that both players should want to play in
the same general area, but the preferred moves often differ, as you
noticed with e.g. the monkey jump.

> We then do an 8 move lookahead using GNU Go's first choice for each 
> player. At the end of the 8 move lookahead we use get_influence to 
> evaluate the board.

How do you transform the influence values into a score? You may want
to take a look at the territorial values computed in the influence
function in 3.4.

> Game 1: W+48.5
> Game 2: W+14.5
> Game 3: B+36.5
> Game 4: W+11.5
> Game 5: W+24.5
> Game 6: B+16.5
> Game 7: W+5.5
> Game 8: W+21.5
> Game 9: W+5.5
> Game 10: B+0.5

Looks okay but of course way too few samples to be statistically
significant. Also an interesting question is how your lookahead
compares to a single GNU Go spending a comparable amount of
computations by playing at a higher level.

> I poked around for a while and did not find the default komi used in 
> the above scoring.

Probably 5.5.

> We have the .sgf files of these games and a logfile containing the 
> evaluation criteria, should anyone have more interest. We are still 
> changing the details, so this set of 10 games will not really compare 
> with future sets.

It would be interesting to see at least some example game. If your
implementation allows it it would also be enlightening to know how it
does on the regressions. (Please ask if you don't know what's needed
to test that.)

/Gunnar