HanabiAI from Facebook achieves state-of-the-art results

2019-12-09

Facebook AI department has announced that it has created the latest AI that can achieve state-of-the-art performance in playing the card game Hanabi. The AI system achieved a score of 24.61 out of 25, surpassing the previous best score of 23.92.


Unlike AI challenges in games like chess or Go, Hanabi is a cooperative game where participants can work together to achieve a common goal.


In a phone interview, Facebook AI researcher Noam Brown said, "The results we've seen far exceed our expectations and those of other researchers. In fact, the gains we've obtained from search are stronger than those obtained by all previous deep reinforcement learning algorithms."Facebook's Hanabi AI draws on some of the search techniques used by Pluribus, an AI developed by Facebook earlier this year that defeated some human champions in poker. The AI team at Facebook achieved this feat by combining deep reinforcement learning with search techniques. The search algorithm transforms the problem into a single-agent setting by having all agents except one execute agreed-upon policies, using a reinforcement learning algorithm called blueprint. According to the paper titled "Improving Policies in Cooperative Partially Observable Games through Search," this blueprint allows the search agent to "treat the known policies of other agents as part of the environment and maintain beliefs about hidden information based on the behavior of others."


Brown believes that ultimately, AI similar to Facebook's Hanabi robot can help robotic systems, autonomous vehicles, or conversational AI agents better respond to human activities by addressing the "theory of mind" challenge. Theory of mind is like putting yourself in someone else's shoes to infer their thoughts and next actions. For example, when you're driving and the car in front of you stops, you might infer that someone is walking towards a pedestrian crossing, even if you can't see that person yet.Brown said, "This is something that humans do very naturally, this idea of being able to put yourself in someone else's shoes and understand why they're taking the actions they're taking, what they're thinking, even if they're not taking action. Not knowing certain things. But this is something that AI has historically struggled with. There's been a long-standing debate about whether non-human primates have theory of mind and at what age human infants develop theory of mind, and I think it's really interesting to see this behavior emerge in artificial intelligence. I think it will be very important if we want to deploy AI in the real world to interact with humans because humans expect this kind of behavior."

RECENT AI NEWS

RECENT AI TOOLS