We all want to be great negotiators, but most of us don’t do it that effectively. How many times have you tried to make a deal, only to blink first and give up the one thing that you cared most about?
A new research project from Facebook aims to help us get more of what we want—and save us the hassle of having to go head-to-head to do so. The project, from Facebook’s artificial intelligence research group (FAIR), builds on the social network giant’s work over the years to enable tools like chatbots to engage in short, natural-language conversations. But while current publicly available bots are capable of doing things like booking a restaurant table, they’ve proven less adept at carrying on a meaningful conversation due to the challenge of understanding what we’re saying and combining that with their knowledge of the world we live in.
The new project—which is not expected to result in a consumer product anytime soon—is meant to demonstrate that it’s possible to build chatbots that can think ahead and plan, said Druhv Batra, a FAIR research scientist visiting from Georgia Tech. “What we’d like these agents to discover,” Batra said, “is that you have to think a few steps ahead to come up with natural-language plans and come up with something that makes everyone happy.”
As with many Facebook research projects, the company is planning on open-sourcing the work.
It’s not yet clear how sophisticated the bots’ negotiating skills will become—and no one should expect them to be working out treaties or anything complex. But in the early going, FAIR imagines that they could be good at scheduling meetings between two busy people, or finding a mutually satisfactory time to go to the movies, or settling on a desirable sale price for some consumer product.
At the core of the project is a brand-new technique that Facebook calls dialogue rollouts, in which bots simulate the direction that a negotiation might go all the way to its end in order to figure out the best possible outcome.
The idea is that both sides would give a score to their own best outcome—via a system in which the bot asks you questions related to the negotiation at hand—for example, what type of cuisine you prefer, what nights you could work late, or when you might make exceptions to an otherwise busy schedule. Then, the two dueling bots would be incentivized to search for the outcome with the maximum score. Fail to reach an agreement, and both sides score nothing. Obviously, a score of zero benefits no one, so the system is built to maximize the chances of the best possible outcome for both sides.
As Mike Lewis, a FAIR research scientist who worked on the project, put it, one of his team’s big challenges was getting bots to learn to compromise, since both sides can’t always get what we most want. So the goal is coming up with a deal that will make both sides at least somewhat happy.
In the project’s early days, it hasn’t kept track of negotiation results. But Lewis said that the FAIR team plans on exploring ways to teach the system to learn over time so that one side can’t be exploited time after time by being forced to accept a less-than-ideal outcome.
It may be working. In a paper on the research, FAIR wrote that over the course of the project, they found that in negotiations between dialogue agents and humans, their bots were increasingly sending the last message less often, “which we can insinuate means they’re proposing the final agreement more often than accepting the counter offers.”
In the meantime, Batra explained, one of the best things to come out of the project has been watching the bots learn good negotiating strategies, including one that many of us probably do understand—initially asking for more than we want and then settling for something closer to our true aim.
“They figure out that one negotiation strategy is to ask for more than what you want,” Batra explained, “and then concede. No human told them that. It emerged naturally.”