Ludii Forum
Wrong heuristics for Misere Bamboo - Printable Version

+- 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: Wrong heuristics for Misere Bamboo (/showthread.php?tid=857)



Wrong heuristics for Misere Bamboo - Michael - 03-03-2022

I think there is a problem with the Bamboo AI. I played a game of Misere Bamboo on size 4. I think I gave AB one minute. And it was roughly using a strategy for standard Bamboo. It seems like the heuristics are not option-specific. I have attached a trial. I played as white (P1).

Also, another thing: I have been wondering what those custom AI things are. In Bamboo it says "Bamboo_ai". Are those agents that use heuristics that some machine learning algorithm has found? In particular, I wonder what happens if the implementation of the game is changed. In Bamboo, for example, I am planning to change the implementation to take advantage of the new group quantifier that wasn't available when I scripted it. Will this make the "Bamboo_ai" useless, so it must be trained again?


RE: Wrong heuristics for Misere Bamboo - vetasnead - 03-29-2022

Did you find answer? Looking for same information. Thanks


RE: Wrong heuristics for Misere Bamboo - DennisSoemers - 04-01-2022

In most cases we indeed currently only have AI metadata (including heuristics) for the default options of a game. Or sometimes specifically for those combinations of options that are also represented by named "Rulesets". Having separate metadata for every combination of options in every game is not feasible given how many different options, sometimes with many different values, some of our games have. We focus only on the rulesets since most of those are relevant to our research in the Digital Ludeme Project.

For some games, the _ai.def files indeed contain heuristcs and/or features that have been optimised automatically through self-play. In Bamboo this indeed seems to be the case (it's generally the case when you see a bunch of different numbers with lots of digits). In other cases, we only (for now) have simpler heuristics found with a simpler search (simply trying several different basic heuristics and evaluating each of them, without further weight training).

If you change a game's rules, it is indeed likely that trained heuristics/features would not work as well (though, depending on how big the changes are, some similar strategies may or may not still apply and the heuristics may or may not still work reasonably well).