I honestly think there are better ways to handle this, if you must, other than changing the goal to a losing goal. At least give the computer the right goal (winning) and adjust from there.
If I were trying to solve this "problem", the solution I would look at first would be to pre-process the moves in advance of the search, and just impose a very slight bias on the root move list. The idea is that all other things being equal, it will play a patterned or "normal" move instead of a random move when the game is virtually over. If you do it right, you can probably make it play more like like it has an ego without making it weaker. - Don Petr Baudis wrote: > On Fri, Mar 07, 2008 at 08:04:37AM -0500, Thomas Wolf wrote: > >> I assume that when you change komi dynamically, all that was learned >> by MC so far under the different komi value is useless/wrong. >> > > But what are actually your reuse rates? With the standard UCB1 formula, > I find reusing branches from earlier trees give relatively insignificant > number of extra playouts, at least with p=0.2 (I find that using lower p > values leads to very deep UCT trees, but the choice of moves being read > out is _very_ noisy). Maybe this is much better with RAVE or such? > > _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
