[Computer-go] MC in games of imperfect information
nick at maproom.co.uk
Fri May 14 08:14:06 PDT 2010
In message <960661D6-7942-4CD5-BB94-8B11518DFF60 at gmx.ch>, Isaac Deutsch
<ibd at gmx.ch> writes
>I'm thinking about creating a computer player for Tichu
>(http://en.wikipedia.org/wiki/Tichu), a game that is rather widespread
>here in Switzerland. However, to my knowledge there exists no official
>bot that plays it. Someone I know has created a bot that plays using
>rules only (if X, play cards YZ), but his findings were that the bot
>plays pretty weak.
>Of course, the game is solvable when there are only 2 players left
>because then, the distribution of cards is clear and the best strategy
>can be calculated.
>With 3 or even all 4 players still in the game, it is clear that it is
>not clear which player has which cards. :) It is a game of imperfect
>information. I was wondering if Monte Carlo (MC) can and should be used
>at this stage of the game. It seems that there has been some success of
>MC in poker.
>If yes, how would MC be used? When a move is to be made, should all
>legal moves be generated, and should a number of playouts be simulated
>for each (1-ply MC)? What do the playouts look like? Give all players
>(weighted) random cards, then play the game out with deterministic
>rules? Give all players (weighted) random cards, then play the game out
>(weighted) randomly? I'm pretty much just trying to think of something
>based on what 'works' in Go. :) Because of the uncertainty in the
>distribution of cards gives so many possible combinations, it seems
>impracticable to create a tree with all possible distributions.
>If no, what alternatives are there? Neural networks?
Tichu is far closer to bridge than to Go, or Poker. There has been some
successful work on programming bridge. I suggest you talk to a bridge
programmer. I don't know any, but there is at least one who follows the
usenet group rec.games.bridge.
Nick Wedd nick at maproom.co.uk
More information about the Computer-go