Artificial intelligence has evolved so much over the years that it has begun to replace humans in many fields of life and this does not only mean mundane tasks like scheduling or organizing things. Robots have now become intelligent enough to be able to beat a human player in games like chess, and the growth is to the extent that Elon Musk predicts that they will be able to take over all jobs in a decade or so.
While Elon Musk’s prediction still sounds a bit far-fetched, it is not entirely untrue. One thing that the artificial intelligence lacks as of now is the ability to negotiate. Humans are natural negotiators, making up things, sometimes lies, flattery or manipulation to coax people to get their own way. New research from Facebook has given rise to an artificial intelligence system that can just that. The AI has learned the art of negotiation by studying 5,808 human conversations, and this could lead us to robots that can schedule meetings on their own, or even get us the best shopping deals online.
Facebook researchers created a game to help the bots learn haggling over books, hats, and basketballs to get the best deal. Each of these objects was assigned a point value, and each of the bot negotiators was to divide them over text conversation. The AI was to see every object’s worth and generate a statement to get it. As the books were worth more points, the bot would say, “I’d like all the books.” Based on what the machine has learned from the human negotiation conversations, it will generate a combination of words that will bring back the highest reward.
Once the neural network has generated the statement, it will anticipate the likely responses of the opponents and then generate a response of its own. This will go on until the conversation ends and the bot narrows down to the best deal it can get. The system was taught not to accept coming empty handed, so it can never get out of the conversation.The pursuit of Facebook’s AI isn’t too different than other applications of AI, like the game Go.
Just like Google AlphaGo, the Facebook neural network anticipates the opponent’s actions and generates optimatal responses. The Facebook AI is, however, different in a way that it needs to make sense to humans while it does the negotiations.
One thing that probably even the researchers did not expect from the neural network to learn was the art of lying. On certain occasions, the AI did not outright say what it wanted, instead, it pretended to have an interest in some worthless object only to use it later to get what it actually wanted. Facebook has not really figured out if the AI learned this from human hagglers or if it discovered the trick on accident. Whichever way it was, the technique worked perfectly. The Facebook researchers say they can control how much the bot can negotiate by varying the AI’s capacity to change its responses.
Knowing Facebook’s growing efforts towards building bots to negotiate on behalf of users on the Messenger platform, this research was expected by most. Facebook announced its dream to make the Messenger a platform for future of customer service, on the 2016 F8 developer conference. While that has not become too close to reality, the rapid growth of AI is noticeable. To make the research more useful and accelerate the work being done to make AI capable of generating useful text and understand querries, Facebook is making all their work in the field open source, including the code and the research data.