+- Ludii Forum (https://ludii.games/forums)
+-- Forum: Suggestions (https://ludii.games/forums/forumdisplay.php?fid=10)
+--- Forum: Submit Your Games (https://ludii.games/forums/forumdisplay.php?fid=23)
+--- Thread: Tumbleweed (/showthread.php?tid=197)
Pages:
1
2
RE: Tumbleweed - DennisSoemers - 10-05-2020
> In the case when the AI is assessing the position at the end of a MCTS play-out, and its play-out has reached a position where it has run out of moves, will it always understand that such a "certainly lost game", is actually certainly lost?
No it won't, because that's not always true. Maybe it is always true for your specific game, but that's domain knowledge which our general game AIs don't have access to.
RE: Tumbleweed - AlekErickson - 10-05-2020
If the AI performs random playouts (until a terminal position) from all positions reached in the MCTS to assess the percentage of wins, then it would find that 100% of positions where it is forced to pass, are lost positions, despite not having any domain-specific knowledge of this game.
RE: Tumbleweed - DennisSoemers - 10-05-2020
Oh yeah you're right of course. It can still affect the AI playing strength though, because it won't **instantly** know to just treat that position as a loss and back-up a loss. There will be a (potentially big) subtree below such a node, which an algorithm like MCTS will spend time in with multiple random playouts, each of which can also take a longer-than-necessary time.