Are there general strategies for games with incomplete information, especially trick-taking games like Bridge and Doppelkopf?
I am interested in ways to implement AI strategies for such games.
The bounty is for the answer with the best description of a particular strategy.
I would like to contribute with some concrete information for the card game Doppelkopf, which the author asked exemplarily. In 2012 a master thesis by Sievers was written where he adopted the UCT algorithm for the Doppelkopf game.
UCT normally assumes a perfect information game, thus he first solves the "card assignment" problem, which is to guess a card assignment for each player based on some already known cards.
After solving this, he tried two ways to perform the algorithm with his solution to the card assignment problem:
1) Guess a card assignment for each UCT-tree and look at the average of multiple trees. He calls this strategy ensemble UCT.
2) Take a single uct tree and guess for each rollout a new assignment. In the selection phase of UCT you simply ignore all inconsistent children. He calls this strategy single UCT.
My feeling is that 2) makes a stronger AI but it seemed to be weaker, which he pointed out more clearly in a follow-up conference paper from 2015.
Inspired by AlphaGo's success, I started a project with a friend for his bachelor thesis and he made a policy neural network using a character based LSTM to guide the selection process of the UCT algorithm. His bachelor thesis only covers some test results for the ensemble-UCT but I already tested it for the single UCT player too and it makes a much stronger AI. I guess this is because the single UCT player benefits alot more from reducing the search space more effectively.
So this answer is more or less the same as @charley have been given but a little more concrete.
