Facebook Teaches "Theory of Mind" to AI
Thu, April 22, 2021

Facebook Teaches "Theory of Mind" to AI

The theory of mind refers to understanding the actions of agents / Photo Credit: Phonlamai Photo (via Shutterstock)


AI systems have been shown to beat humanity’s best offerings in competitive games, said Andrew Tarantola of Endgadget, a technology news website. However, real life is not a zero-sum game like poker or Starcraft. Hence, we need AI not to be against humans, but to collaborate with them. With that, researchers from Facebook taught an AI to play the cooperative card game Hanabi to make it understand how humans think. In particular, the researchers instilled the AI model with the theory of mind. 

Noam Brown, a researcher at Facebook AI, informed Engadget,  "Theory of mind is this idea of understanding the beliefs and intentions of other agents or other players or humans." The theory of mind is something that we developed from a very young age,” but it is something that AIs struggled with for a long time. Brown explained that the AI is trying to put itself in other players’ shoes and ask why they take certain actions to infer about the “state of the world that it can’t directly observe.” 

Hanabi was created by French game designer Antoine Bauza in 2010. Two to five players are tasked to construct five, 5-stacked cards. The stacks are color-coded and must be arranged numerically from one to five. Players must complete all the stacks or try to get as close to scoring 25 points once the team has run out of moves. None of the players know what’s in their hands, as the cards are facing away from themselves. Thus, their teammates know what they hold.  Players can tell their teammates either the color or number of cards in their hands. Overall, Hanabi pushes players to figure out the rationale behind their actions. 

Facebook’s Hanabi AI creates a rough “blueprint” of what could happen in the game and leverages that information to formulate a “ near-optimal strategy in real-time based on what cards are currently in play.” Moreover, its single-agent search offers “enough of a predictive boost” to put AI ahead of even the best Hanabi players, while multi-agent search helps it earn near-perfect 25 score points. 

Obtaining near-perfect scores in Hanabi sounds great, but Facebook has plans for its cooperative AI. Lerer explained, “What we're looking at is artificial agents that can reason better about cooperative interactions with humans and chatbots that can reason about why the person they're chatting with said the thing they did.”