+- Ludii Forum (https://ludii.games/forums)
+-- Forum: Problems (https://ludii.games/forums/forumdisplay.php?fid=5)
+--- Forum: AI Problems (https://ludii.games/forums/forumdisplay.php?fid=7)
+--- Thread: Yavalath (/showthread.php?tid=137)
Yavalath - RogerCooper - 08-17-2020
I was playing Yavalath and I see that Luii prefers Biased MCTS. However, when I match it up against Alpha-Beta, Alpha-Beta is superior.
RE: Yavalath - DennisSoemers - 08-17-2020
Hmmm that's interesting.
I've just run some games on my own machine, and that seems to confirm your statement; I don't have statistical significance yet, and it's not 100% wins for AlphaBeta, but indeed AlphaBeta does seem to be winning more.
When I run the evaluation games for all games, I don't do so on my personal machine, but on a compute cluster (of course we have way too many games to do this all on my own machine). That was last done shortly before the 1.0.0 release, and then Biased MCTS (specifically the variant with uniform playouts) did come out on top, even against AlphaBeta. This could be due to hardware differences... but I'm not sure what the best way to solve it would be if the difference is indeed due to that. Different people are going to have different hardware, and this can change which algorithm performs best (as can the time control settings).
Note that not all games have completely transitive relations between AI playing strength. For example, it is possible to have a game where AlphaBeta beats Biased MCTS when matched up against each other, but where Biased MCTS performs better than AlphaBeta does against all our other AIs. In such a case, we would still prefer Biased MCTS as the more "robust" agent. Though, looking at the cluster results, this is not what seems to be going on in this particular case, since on the cluster Biased MCTS also beat AlphaBeta in that specific matchup...
RE: Yavalath - RogerCooper - 08-18-2020
Perhaps the problem is that in Yavalanth, there are separate win and loss conditions with the conditions being similar. This type of situation may confuse MCTS, as it would trouble distinguishing between good and bad lines.