Facebook Is Training Its Artificial Intelligence to Negotiate
The company’s artificial intelligence is learning how to bargain. In some cases, it’s already just as good as humans.
PHOTO CREDIT: Getty Images
Now the social network's algorithms know how to negotiate. In a study published Wednesday, Facebook revealed that it has taught chat bots to bargain with humans--and that they did so successfully. So much so, that Facebook claims the human opponents often didn't realize they were interacting with bots.
Facebook A.I. Research (FAIR) used transcripts of negotiations between real people to teach the bots to imitate their responses. FAIR then had them practice negotiations against fellow bots, rewarding them for positive outcomes and thus reinforcing words and behaviors that led to those outcomes. Over time, the bots learned to go beyond simply mimicking humans and instead became more unpredictable with their responses.
To test the model's effectiveness, Facebook created scenarios with a hypothetical set of objects. The A.I. agent and the human were given different point values for the various objects. By communicating back and forth, they both had to agree on how to divvy up the objects. If they couldn't reach an agreement within 10 back-and-forth exchanges, both participants received zero points. Facebook arranged the trials so that it was impossible for the two parties to divide the points equally.
As in most real-life negotiations, the participants didn't know how the other one valued the specific objects and instead had to infer it from their dialogue. To this end, the bots learned how to bluff, feigning a desire for objects that weren't actually important to them while downplaying their desire for those that were.
This level of planning ahead paid off: Facebook says the bots won the negotiation just as often as the humans did.
Earlier this year, the Libratus A.I. system, created by a team of Carnegie Mellon researchers, handily defeated some of the world's best poker players. That computer relied on bluffing tactics that were once only achievable by humans.
The designer of that system, Tuomas Sandholm, told Inc. it could one day be useful in business settings, advising companies on which deals to pursue and which to leave on the table, or aid in salary negotiations.
Of course, some see potential dangers in giving A.I. that kind of power. In April, Tim Berners-Lee, inventor of the world wide web, laid out the possibility that business-focused A.I. systems could someday run companies without needing human intervention. They could then spawn new companies, he said, and eventually would control the institutions that compose much of the world's economy.
"You have survival of the fittest going on between these A.I. companies," he said at the Innovate Global Finance Summit, "until you reach the point where you wonder if it becomes possible to understand how to ensure they are being fair--and how do you describe to a computer what that means, anyway?"
Facebook used an "end-to-end" training model, which means the process could be altered to give the algorithm other, similar goals to the one in the study. The company also open sourced the code on Github so anyone can tinker with it. Conceivably, it could be trained to negotiate things like meeting times or--thinking outside of a business setting--fantasy football trades.
In an email, Dhruv Batra, a Facebook visiting researcher who worked on the project and also teaches computer science at the Georgia Institute of Technology, told Inc. that Facebook doesn't have any plans to implement the technology into its product yet. Instead, Batra sees it eventually being used to negotiate things like car or house prices--"basically," he said, "any scenario where humans may want to offload the unpleasantness or mundaneness of semi-cooperative dialogue to an agent."