Hi Francois, > You mean all the 3x3 patterns? I'm only using 3x3 > patterns that occur a number of times in my training > collection.
If 3x3 is the only pattern size you use, that probably does not apply. But if your pattern system supports other sizes then you must have implemented some kind of "not found" mechanism. All not found patterns will have identical urgency, the urgency of "not found". As the number of 3x3 patterns is small, < 64K patterns because there are ill-formed patterns you cannot find, like patterns where "off the board" points do no form a side or two sides sharing a corner. So my advice is: regardless of if you found them or not, include all legal patterns, i.e., all patterns where "off the board" points are as in (1,1), (2,1) or (2,2). In playouts, sooner or later you will find them all. > I'm not following you here. What sort of classes? Degrees > of liberty? R^40? Would you mind explaining a bit for me? Once you remove the ill-formed patterns and even considering mirror and rotation, you still have tenths of thousands of patterns. Each one has the right to have its own gamma value. But you cannot reasonably optimize so many gamma values, therefore the idea is to group them. Some categories are just standard go relations as you can see in the Mogo papers "threatens a cut", "bad keima", "contact move" etc. Other are related with the height (isolated stone in 1st row vs isolated stone at least on the 2nd) and the ones I have tried most are "life preserving" For instance in a 3x1 shape. OOOOOOO OXXXXXO OXabaXO ------- Clearly, a is much worse than b for both players. a for X kills its own group before it has a 1-point eye. a for O does nothing good, probably forces X to play at b and save the group. So it is clear that the shape of b should have more urgency than that of a. Other shapes are related with connecting your stones, and finally there are classes where you put all the patterns that do not fit in any of the previous but may be, on a side, on a corner, with 1 liberty, with 2 liberties etc. My classification was done by trying to make the most of 3x3 patterns and to identify as a class all the ideas I found in literature plus the "life preserving" ideas. But I cannot say it works as good as I would like. Something better can be done for sure. When you have done such a classification, all the patterns in the same class have the same gamma value. The gamma of the class. With 40 classes (in my case) you "only" have to tune 40 gamma values, not all at a time. Say you let 8 change randomly (x.25, x.5, x1, x2, x4) and keep the other 32 constant. Then you see the result of around 2000 games and find 1, 2 or maximum 3 of these values working consistently better high than low (or vice versa) across the 2000 games. These are the candidates to be changed. The initial set of gammas is uniformly random (all equal). But please, feel free to find better ideas and share them ;-) Jacques. _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
