[Computer-go] effectiveness of transposition tables for go
tesujisoftware at gmail.com
Thu May 6 12:33:27 PDT 2010
During my brief stint in playing with MCTS I also tried to implement
transpositions. I read somewhere (about MoGo I think) that it would
gain about 100 ELO points. But I didn't see any gain at all. David
Fotland told me he used it and it did work for him.
Possibly my idea how to do it were plain wrong. I never had time to
look into it very carefully. What I tried to do was creating a
hash-table next to the tree. The hash-table contained lists of
positions that were identical but with different positions in the
tree. So a hash-table of lists. When updating the win-rate in the
tree, I wouldn't just go up the one path, but up multiple paths when
there were transpositions (taking care of loops). Basically the tree
had become a graph.
What I saw was a small gain for small number of nodes ( a few K), and
worse results the larger the number of nodes became to the point it
was a negative contribution.
More information about the Computer-go