On Sun, 2007-03-25 at 15:22 -0400, John Tromp wrote: > On 3/19/07, Don Dailey <[EMAIL PROTECTED]> wrote: > > I'm testing a future Anchor player for CGOS. I am calling > > it FAT for Future Anchor Test! > > > > It plays fixed depth and I pre-calculated what level to make > > it play at 1800 strength. I came pretty close, Fat-25 is > > playing at 1836 at the moment and doesn't require too much > > CPU power. It's Lazarus scaled down to play fast. > > Isn't AnchorMan a program that plays pure MC with uniformly random > (unbiased except for eyeflll) moves?
Yes, exactly. AnchorMan does exactly 5000 play-outs before making a move. When I say the level is fixed - what I really meant was that it is set to 5000 and is not based on the clock or the power of the computer it runs on. I can't say AnchorMan does nothing fancy though. It's not "pure" because there are minor bonuses and penalties "superimposed" on certain moves in order to encourage/discourage correct play. In my testing it is worth quite a bit. For instance, if a move is self-atari and most of the play-outs indicate that the point under consideration will end up belong to the opponent, I have a fairly strong anti-move incentive. I apply a fixed penalty to the score that move would normally recieve. > Then it would be good to have as 2nd anchor a program that plays > pure UCT with pure MC and nothing fancy. I have a version of Lazarus that represents such a creature. I think it plays about 1700-1720 on CGOS. The play-outs are unformly random except for eye-filling moves and it is UCT. I could even publish the algorithm in pseduo code so that it could be duplicated exactly without ambiguity. However, to play that strong it would not be a low resource program. That's why I constructed Fat to play at 1800 and it can make the time-control on a fairly slow computer (it takes a total of about 1 minute of time on my core 2 duo.) The problem is where to run an Anchor? I don't want to dedicate my own computer to running an Anchor 24/7. The current Anchor on CGOS actually runs on the same machine that is the server. But it sucks very few cpu cycles. A stronger Anchor, even FAT uses quite a bit more resources - and I want to be a good boardspace citizen - I won't run a strong Anchor directly on the server machine. > If parameters like #simulations and exploration coefficient are fixed, > then people can test their own implementations by seeing if they can > closely match this UCT anchor in rating. Yes, it's a good way to test your implementation and a good idea. > I'm assuming here that it wouldn't take that much computing power to > make such a program play at an 1800 rating. Fat is playing about 1800 and as I mentioned, it doesn't take much computing power but it uses heavy play-outs. That makes it fairly complicated to describe. A 1700 version of Lazarus with the lighter simpler algorithm is easy to describe but uses more resources. In either case, I will have to get someone as a volunteer to run this Anchor and my idea is to have 2 or 3 copies so that no single person has to run it continuously. I am willing to consider other anchor possibilites, if someone want to donate their binary and thinks it would be a good reference program. - Don > regards, > -John _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
