For 9x9 games, when I added progressive widening to AntiGo (before I added RAVE), it was low hanging fruit. I used my old program Antbot9x9 for the move ranking and got a very nice strength increase for very little effort. Then, with a bit of tweaking, I got more improvement. RAVE, on the other hand, gives me a small improvement, and it comes at the cost of having to fuss with it endlessly.
With no opening table and 5000 playouts per move, AntiGo wins against gnugo3.7.10 in the neighborhood of 60 and 70%, depending on what is enabled. I optimized my playout policy before I added RAVE. They seem to interact poorly. The following tables are from running UCT for black to move on an empty 9x9 board and showing the RAVE scores at the root. Here is the result when using light playouts. RAVE ?? 1|?? 33?? 43?? 46?? 38?? 40?? 41?? 28?? 33?? 41 ?? 2|?? 37?? 44?? 37?? 35?? 43?? 42?? 37?? 42?? 40 ?? 3|?? 42?? 44?? 41?? 43?? 45?? 48?? 44?? 44?? 39 ?? 4|?? 35?? 45?? 46?? 45?? 43?? 42?? 48?? 45?? 35 ?? 5|?? 36?? 42?? 43?? 45?? 45?? 44?? 40?? 45?? 36 ?? 6|?? 40?? 42?? 40?? 43?? 45?? 43?? 45?? 43?? 37 ?? 7|?? 41?? 47?? 45?? 38?? 42?? 38?? 39?? 40?? 35 ?? 8|?? 32?? 45?? 44?? 45?? 46?? 47?? 45?? 40?? 38 ?? 9|?? 38?? 40?? 38?? 27?? 40?? 38?? 42?? 30?? 37?? It looks noisy, but perfectly reasonable. With more playouts, it looks nicer. Here is the result when using my heavy playouts. RAVE ?? 1|?? 53?? 44?? 45?? 44?? 45?? 47?? 46?? 43?? 54 ?? 2|?? 43?? 50?? 49?? 48?? 49?? 49?? 50?? 50?? 44 ?? 3|?? 46?? 49?? 49?? 50?? 49?? 49?? 48?? 49?? 46 ?? 4|?? 47?? 48?? 48?? 47?? 49?? 47?? 48?? 49?? 47 ?? 5|?? 45?? 49?? 47?? 49?? 49?? 50?? 47?? 49?? 47 ?? 6|?? 46?? 48?? 48?? 49?? 48?? 48?? 49?? 49?? 45 ?? 7|?? 47?? 50?? 49?? 50?? 49?? 48?? 49?? 50?? 45 ?? 8|?? 47?? 49?? 50?? 50?? 48?? 48?? 49?? 51?? 44 ?? 9|?? 54?? 45?? 47?? 45?? 46?? 46?? 46?? 46?? 54 It looks like garbage! In fact, it's surprising that it could help at all. Progressive widening saves that day. While moves in the corner have higher RAVE scores, they have very low rank. Here is the result when using my heavy playouts and CRAVE. CRAVE ?? 1|?? 35?? 37?? 37?? 28?? 40?? 35?? 40?? 32?? 25 ?? 2|?? 38?? 35?? 44?? 39?? 38?? 49?? 39?? 42?? 34 ?? 3|?? 33?? 38?? 48?? 48?? 51?? 49?? 49?? 39?? 36 ?? 4|?? 41?? 41?? 48?? 50?? 50?? 47?? 49?? 41?? 37 ?? 5|?? 35?? 41?? 51?? 48?? 50?? 47?? 50?? 43?? 31 ?? 6|?? 39?? 40?? 51?? 50?? 50?? 47?? 51?? 46?? 41 ?? 7|?? 36?? 36?? 51?? 51?? 45?? 52?? 50?? 38?? 37 ?? 8|?? 37?? 40?? 45?? 41?? 47?? 37?? 43?? 39?? 32 ?? 9|?? 21?? 27?? 31?? 36?? 38?? 35?? 32?? 35?? 28 Maybe CRAVE (and perhaps, progressive widening on 9x9) seems more useful to me because of compensation for problems other programs don't share. - Dave Hillis
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
