Training AI by using games which learn though player’s behavior and tries to mimic the human problem solving skills, is not a new thing as many games like Go and StarCraft are already trying to do the same. However Google’s Artificial Intelligence research team, Google Brain along with other Google subsidiary DeepMind has released an AI learning environment for Hanabi, a more complex and collaborative card game.
People familiar with the Hanabi card game point out the deceptive complexity of the game as it involves two to five players indulging in a cooperative gameplay setting. The players have to be conscious not about one’s own motives but also the possible intentions of the opponents in the game. The game requires higher levels of reasoning and theory of mind which is an understanding of the mental states of others and the probable differences of mental states between different individuals. The game being one of the trickiest is an intense exercise in cooperation, inference and strategy.
In the Hanabi game which won the Spiel des Jahres (excellence in game design award for card and board games), five cards are dealt to the players from the hand and each player can see cards of other players while their own cards remain concealed form them. The information sharing in Hanabi is limited by the allowed number of hints and the quantity of information that can be communicated to other players though player’s own actions. A player has to selectively reveal information to the other players to win the game, which is a complex situation to understand for AI but if incorporated can enable certain foundational skills which humans use in their social interactions.
Additionally it has to learn to reveal maximum information to other players for helping them succeed. Researchers believe that if the AI agent can be trained to navigate such a convoluted information environment, it will be a huge step towards AI’s effective cooperation with humans. The challenges are equally daring for researchers as it involves bringing together algorithmic advancements across several AI subfields including game theory, emergent communication and reinforcement learning. Emergent communication is the subfield which studies how multiple AI agents communicate in a collaborative setting.
Google’s AI research team tested all the prevalent state-of-the art reinforcement learning algorithms and found them to perform poorly. With the release of open sourced Hanabi environment, Google hopes to give impetus to further work in the AI research community.