Jacques Basaldúa wrote: > Don Dailey wrote: > >> You can use Zobrist hashing for maintaining all 8 keys incrementally, >> but you probably need a fairly good reason to do so. Incrementally >> updating of 1 key is almost free, but 8 might be noticeable if you are >> doing it inside a tree search or play-outs. > > Yes. Don is right. Of course that is not part of the real program, but > of a program that searches the book. In my case (19x19 only) I play a > maximum of 20 moves (10 per player) from the book and then switch to > the "real" program. > > I never shared the naif idea that some openings (played by high dan) are > better than others and that finding a correlation between a given move > and the result of the game was meaningful. I consider all "popular" > openings equally balanced and playable. Finding a move in the book > just saves you the time of 4-5 moves (10 if you are really lucky), gives > you a straightforward way to randomize the opening (drawing between all > moves in the book uniformly) and makes the board contain some > information when the real thing starts. > Indeed, my book scheme for Lazarus is very simple. I track the first move out of book for Lazarus and deep search it N times (for variety.) The next time Lazarus encounters that same position, there is a book response and Lazarus saves search time. Lazarus plays moves in same proportion they are selected in the "deep search" process. In the opening position, if e5 was selected 5 times and d5 1 times, Lazarus would play e5 5/6th of the time.
In all games, Lazarus writes to a file the first move out of book and it is placed in a queue of moves to be deep-searched. In practice, the book search is slower than on-line play but this could be adjusted. I'm building up moves to be searched quicker than I am searching them. I could solve this by setting N smaller and/or searching faster but I prefer nice deep searches with reasonable variety. I think N is 4 right now. This doesn't guarantee that the book moves are high quality, but it does have the desirable features that the search is better than during a real game and it saves time. I suspect saving time for future searches is more important than the improved quality of the opening moves. Unfortunately, if Lazarus improves I have to rebuild from scratch (unless the improvement was very minor.) Also, the book would not be useful for games at higher time-controls than the book was searched at. - Don > Jacques. > > > > _______________________________________________ > computer-go mailing list > [email protected] > http://www.computer-go.org/mailman/listinfo/computer-go/ > _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
