Outfoxed by a bot? Facebook is teaching AI to negotiate

One day, the art of the deal might just involve letting artificial intelligence do your dirty work for you.

Researchers from Facebook Artificial Intelligence Research (FAIR) have created AI models, or what they call dialog agents, that can negotiate, according to a blog post Wednesday. They're publishing open-source code as well as research on those dialog agents, the result of about six months' work on the project.

The idea is that negotiation is a basic part of life whether you're picking a restaurant with friends or deciding on a movie to watch. But current chat bots aren't capable of much complexity. Their state of the art is to do simple tasks like book a restaurant or have short conversations of limited scope.

FAIR worked on the problem of how to get dialog agents to operate like people -- that is, come into a situation with different goals and eventually reach a compromise.

The effort is part of a broader push by Facebook to get us to use chat bots. At its developer conference in 2016, founder and CEO Mark Zuckerberg walked through scenarios in which you might use a bot to interact with a business, for example, to order a product or get customer service help. While tech giants like Facebook, Google and Apple are keen to build the personal digital assistant of the future, today's helpers still lack the necessary skills.

It's just one stitch in the larger fabric of work by Silicon Valley, academic researchers and the business community in the area of artificial intelligence, driven by powerful chips, fast networks and access to massive amounts of data about how people lead their digital lives. That's showing up in everything from sorting photos on Facebook to beating Go champions and diagnosing medical conditions.

FAIR didn't delve too far into what applications might be appropriate for bot-bargaining or whether this capability will surface in any Facebook products. But the post did mention this could be an advantage for bot developers working on chat bots with the ability to "reason, converse and negotiate, all key steps toward building a personalized digital assistant."

Negotiation, the FAIR post explains, is both a linguistic and reasoning problem. In other words, you've got to know what you want several steps down the road and be able to communicate it.

In one example, dialog agents were tasked with dividing up a collection of items like five books, three hats and two balls. Each agent had different priorities and each item carried a different value for each agent. The AIs were taught, in a sense, that walking away from the negotiation wasn't an option.

The ability to think ahead is crucial and with the introduction of something called dialog rollouts, which simulate future conversations, the bots were able to do so.

Or as FAIR scientist Mike Lewis put it: "If I say this, you might say that, and then I'll say something else." Lewis said those rollouts are the key innovation in this project.

The research has boosted performance in using various negotiation tactics, like being able to negotiate until there's a successful outcome, propose more final deals and produce novel sentences. The agents even started pretending to be interested in an item so they could later concede it as if it were a compromise.

Humans had a chance to try out the agents, and the researchers said the people couldn't tell they were chatting with bots.

CNET Magazine: Check out a sample of the stories in CNET's newsstand edition.

Batteries Not Included: The CNET team reminds us why tech is cool.