Jonas Kahn wrote: > > This looks very much like the way human players work (albeit with a > tree): read local sequences and outcomes that can be kept in reserve for > a long time, but called about any time depending on the situation. Big > chunks. > > I had the idea that a tree could be added later if needed, but that this might replace a tree entirely as it would be a kind of "meta-tree", or a higher level tree if that makes any sense.
Also, if you look at the annotation of chess games (it's presumably the same for GO but I haven't studied go) you will see something like this: if 5. c3 then 5 ... d5 6. any 7. c5! etc. or you may see a list of moves to follow up with, ignoring the exact details and timing of them when it's presumably straightforward. I also though of persevering tree-like sequences which is what you seem to be implying. A LOT of sequences are good and there can be some flexibility regarding their timing, especially if they are a response to something an opponent might do. There are a lot of moves that are good ONLY as a response. It's a waste of time to play them until it's necessary. This is one serious limitation of all-moves-as-first for instance. So the mechanism I'm looking for is a kind of all-moves-as-first except that it's much more about the actual situation. By the way, I tried all-moves-as-first with individual 3x3 patterns so that the move signature was not just the point you wanted to play to, but the pattern around it and it failed miserably. One of the early papers on MC-like algorithms was based on using simulated annealing for evolving a strategy for playing the game and thus finding a move. I keep wondering if it's possible to do this in a much more sophisticated way. The method used in this paper basically evolved an ordering for the 81 points of the goban (which specified the order the moves should be played in a play-out phase) but I think there must be richer ways to represent a playing strategy. Subsequent papers on MC determined that a purely statistical approach seemed to be just as effective and from these beginnings we got where we are now. But we stopped exploring the machine learning approaches right away. Perhaps that was best, but perhaps not? - Don > Jonas > _______________________________________________ > 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/
