terry mcintyre wrote: > I've been thinking a bit about the collection of > patterns from games, whether of professionals or of > programs. > > It is possible to get some remarkably high correlation > between the moves played by pros and a predictor - yet > still not have a good program. Why? >
In my opinion it's because playing go is more about the quality of your weakest moves, it's not the average quality of the moves. You pretty much have to play every move at a given level to actually be playing at that level. Of course that's a simplification but gets the point across. Some moves can be found by relatively weak players and the difference between good and great players is in a small percentage of the moves. > One possible answer is that many moves are considered > but never played; this information is not captured by > looking at game records alone. Ordinarily, both > players analyze parts of the game - life-and-death > situations, for example - and know exactly what > outcome to expect. For instance, "the L group is dead" > - therefore, one would not create such a provably dead > shape. > > The game record will show the results of decisions > made by pros, but not the process of rejecting bad > shapes. > > A tree search based only on game records is unlikely > to have enough information to weed out situations > which are almost right - "just a little bit dead." > > Suppose a group can be defended - four liberties in a > row, for example. If the opponent plays inside those > four liberties, you play to divide the area into two > eyes - unless the situation is such that the group has > a second eye elsewhere. Game records won't show such > frivolous plays, but it is essential to know how to > respond to programs which do make such plays. > > It might be worthwhile for tree search to include > patterns which have been generated by life-and-death > solvers, determining the status of groups using moves > which seldom appear in game records, but which are > essential to gather tactical information about the > status of groups, used to make top-level strategic > decisions. > > To summarize, the tree search needs to know about > patterns which are unlikely to ever be expressed in > the game record itself. > > I remember we did talk about this is a group. We came to the conclusion that relevant patterns and positions rarely or never occur on the actual playing board, but are just understood. In a trivial way I discovered this when I was looking for king and pawn endings in a huge grandmaster database. I found very few. And yet these are pretty much fundamental to understanding the game of chess and even influence the openings you play and the goals of the game. It turns out that Grandmasters usually know the outcome of any king and pawn ending that is threatened over the board, and so they basically work around them. And yet they are of fundamental importance in the outcome. In some of my pattern learning experiments, I discovered that only a very small subset of possible patterns occur on the real board, and yet for a game tree searcher it would be pretty important to understand those patterns that are "constantly avoided" in order to understand why they are being avoided. That's why I believe that patterns culled from games are pretty much useless. That probably extends to most learning based on observing games. Nevertheless, I got a pretty good improvement using patterns that Lazarus "never" plays as a form of selectivity in Lazarus. So that is evidence that refutes my idea to a certain extent. I figured if Lazarus would never (or almost never given many opportunities) play to create a certain pattern, it was reasonable evidence that it is not a good move. - Don > > Terry McIntyre <[EMAIL PROTECTED]> > > “Wherever is found what is called a paternal government, there is found state > education. It has been discovered that the best way to insure implicit > obedience is to commence tyranny in the nursery.” > > Benjamin Disraeli, Speech in the House of Commons [June 15, 1874] > > > > ____________________________________________________________________________________ > Be a better friend, newshound, and > know-it-all with Yahoo! Mobile. Try it now. > http://mobile.yahoo.com/;_ylt=Ahu06i62sR8HDtDypao8Wcj9tAcJ > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ > > _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
