Claus Reinke wrote: >> Just out of curiosity, what did you expect from the playouts? > > Nothing in particular, really; at this point I'm just trying to > build up an intuition about what I can or cannot expect from them. > At first, I thought light playouts would not fully explore, but at > least randomly cover all possible games, so we wouldn't get good > move sequences directly, but lots of useful statistics to work with; > while heavy playouts would try to mimick more realistic games, so > the statistics would be biased, but there'd be a better chance of > getting good moves directly. That turned out to be somewhat naive on > all counts, as the answers to my earlier messages showed!-)
> There is no chance of covering all possible games, That's a major understatement! The number of all possible games is really enormously huge. > so uniformly random play (light playouts) definitely ignores some, > and if realistic games are unlikely, and good move sequences even > more so, then they get the least coverage. So, in the presence of > limited resources, "no bias" is a bias on its own. Uniform simulations are badly biased but I wouldn't say it has anything to do with coverage. > And the way eyes and other rules knowledge are encoded for > efficiency reasons introduces other biases as well. Eye rules are not primarily there for efficiency reasons but to save information about the original position from dissipating completely. > Heavy playouts never drop possible sequences, Heavy playouts may or may not drop some moves completely, depending on the design. > they just try to focus the available resources away from sequences > that appear less promising. That can have unpredictable effects, so > the "unpromising" sequences are kept in the pool, at low priority, > and the resulting focussed playouts need heavy testing to get some > measure of confidence in any improvements. No, heavy playouts are not about finding promising sequences. The purpose is to get simulations that more often let living groups stay alive, let dead groups stay dead, let solid territory remain territory, and so on, so that the score of the final position correlates reasonably well with the value of the initial position. That's why uniform simulations are badly biased; solidly connected stones almost always win against more loosely connected but perfectly sound formations, causing the program to play very heavy and inefficiently. > In sum, there doesn't seem to be a good basis for understanding playouts > and their optimisation, other than by trial and error. [...] Some things are possible to understand, such as the "often play close to the previous move" idea resulting in fewer tenukis and therefore fewer opportunities for solidly living groups to die, or for hopelessly dead groups to spring alive. If this is not obvious I would recommend playing a number of games against a uniform random player with some 40-50 handicap stones on 9x9. But as far as fine tuning of playouts go I would say it's pretty much black magic. Actually that's true for everything but the most rudimentary principles. /Gunnar _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
