Rules for the game of hanabi can be found here.
Develop an AI capable of playing Hanabi is not trivial: the game is cooperative, not zero-sum and with imperfect information.
We did lots of research in order to implement the AI, but most of the algorithm is an implementation of the paper Re-determinizing Information Set Monte Carlo Tree Search in Hanabi
Playing against himself the AI is capable of obtaining an average score of 18/25 (that is quite remarkable compared to others AI which don't use deep reinforcement learning). Even against other AI we observed that is capable of adapting quite well (being a cooperative game, the score strongly depend on the capability of all the players).