Facebook AI Learns “Theory of Mind”

Competitively, artificial intelligence is far better than human players, especially when it comes to making split-second decisions and being efficient with time usage. However, we don’t want to compete with AI. Instead, we would like to use their enhanced abilities to help us perform tasks and draw conclusions. However, AI has historically had a hard time figuring out how humans think. A research team from Facebook taught an AI to play the cooperative card game Hanabi (Japanese for “fireworks”) to help it understand how human beings think.

The Theory of Mind for Human Beings

Theory of mind can be defined as being able to understand the beliefs and intentions of other agents within a set framework. Humans learn the elements of motivation and how it affects people’s actions from a very early age. The AI is attempting to learn how humans think by putting itself in their positions and asking why they do what they do. In doing so, it attempts to infer something that it can’t discern from the available information.

Using Hanabi to Teach Inference

Hanabi asks players to construct five stacks of five cards apiece. The piles have to be numerically ordered from one to five, and the team only has a finite number of moves to do so. The catch is that none of the players knows what’s in their own hands. They only know the cards the other players have in theirs. Players are allowed to share information, but they need to pay an “information token” and only have a limited amount of those. As a result, each player needs to determine what’s in their hand based on information from other players.

The AI uses a framework that gives it a set of “rules” that operate within the ruleset of the game. It assumes all other players are using the same guidelines and figures out (by way of a search) the most efficient method of doing things. If something happens that it doesn’t expect, it recalculates its next move. This theory of mind is the first step towards teaching AI the power of “seen-or-implied” thinking where humans’ actions aren’t explicitly connected to motives. However, those motives may be discerned, based on how the individual reacts in a certain situation.