Notes - Artificial Intelligence MT24, Imperfect information games
(Theodore Rombouts, The Card Players)
Flashcards
In games of imperfect information, such as card games where the opponent’s initial cards are unknown, @define what is meant by averaging over clairvoyance.
Performing expectiminimax with only one chance node at the start of the game, averaging over all possible initial configurations.
One approach for playing games of imperfect information is averaging over clairvoyance. This is where expectiminimax is performed, with only one chance node at the start of the game which averages over all possible initial configurations:
\[\max_{\text{a} \text{ action}} \sum_{s \text{ state}} \mathbb P(s) \cdot \mathbf{Minimax}(\mathbf{Result}(s, a))\]
Why is this not optimal?
It wrongly assumes the game is fully observable after the initial state, which leads to problems like:
- Making moves that would guarantee a tie rather than possibly lost
- Making moves which gain information
What is an information set of a player in the context of playing imperfect information games?
A set of game nodes that cannot be distinguished by this player given their current knowledge.