[Computer-go] effectiveness of transposition tables for go
tesujisoftware at gmail.com
Sun May 9 11:22:37 PDT 2010
On May 8, 2010, at 6:46 PM, Martin Mueller wrote:
> Fuego only has a single fixed-size array that stores the tree. It does not use transpositions. Markus tried several variations early on in the development of Fuego, but did not find anything that worked better than a plain tree.
So nobody has made it work yet? That makes me feel a little better, I thought I was the only dummy who hadn't managed :)
Although I didn't spend much time on it, I was pretty convinced that I implemented correctly what I intended it to do. Just the result was disappointing, so I kept wondering if I had missed something really obvious.
I am sure I've seen a reference about it working though. It was an article that summed up all the different pieces in MCTS with their respective gains when compared to GnuGo. It went from a winning percentage of less than 50% to over 90%, IIRC. Anyone remember what the article was?
Of course it's also possible that transpositions worked at some stage and stopped contributing after adding another improvement. MCTS is full of non-intuitive surprises like that, I've found.
More information about the Computer-go