Matthew Woodcraft wrote:Greg Schmidt wrote: > When assigning the credit value for a node which is a transposition, are > *all* parents which point to that node credited, or just the particular one > which led to that transposition in the current continuation? If the former > case, that's seems like the TT would lead to more efficient learning (a bit > "RAVE like" since a single trial results in multiple updates) although > wouldn't it be unwieldy for a node to have a list of pointers back to its > multiple parents?
[It's surely wrong to propagate the value of a playout up via all parents in the same way that you would via the single parent if you didn't have a transposition table. That's because the normal MCTS algorithm relies on the fact that 'most' information in each node has come from the 'most promising' child, and you'd break that assumption.]Well the same thing can be said for RAVE, right? RAVE updates multiple parents from a single result, introduces some biases, yet seems to produce useful results.I'm mainly curious to know if anyone's actually tried using a TT this way (as Kocsis suggests), and if so, what the results were.-- Greg There is a safe way to do a little bit of propagation to alternative parents, but it looks a bit marginal. There are some more details (and a link to a paper) in this old thread: http://hosting.midvalleyhosting.com/pipermail/computer-go/2009-March/017901.html -M-
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
