The hash table contains a linked list of nodes with the same hash index. Your description is pretty close, but I don't remember the exact details. Yes, I keep around some nodes for while, but eventually old nodes go into the past and can be recovered.
I actually have fewer total nodes than you do. The commercial version allocates 30K nodes per CPU core. The version in the world championship had much more, but the commercial version can't be that greedy for memory. David > > > I give up when there are no nodes that area available to be > > overwritten. > > I'm still a little confused. > > First, does the hash table contains pointers into a node pool, or do > you use the node pool as the hash table? > > Second, how and when do you determine which nodes (if any) are > available for overwriting, without doing some sort of traversal? > > Is the following a correct summary of what you do when it's time to > create a new node? > > Go to the one node that would be used for this Zobrist hash. If the > same hash is stored there, use the existing node. If not, either > overwrite this node (if this node is available for overwriting) or not > create a new node (otherwise). > > Also, what exactly is the criterion for overwriting? You said "nodes > that have few visits or are old". Is this just (visits < > visitThreshold) || (depthFromBeginningOfGame < currentGameTurn)? > Doesn't depthFromBeginningOfGame cause you to keep around nodes that > are outside the still-relevant branch of the tree (i.e., the branch > descending from the current root), almost all of which are irrelevant? > > > If the search is long enough, there are no nodes that can be > > recovered, and the > > UCT tree stops growing. I keep searching when there are no reusable > > nodes. > > > ...because this improves the quality of information in the nodes you > were able to create. That makes sense. > > Thanks, > > Peter Drake > http://www.lclark.edu/~drake/ > > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
