Subscribe now

Technology

Chatbots learn how to negotiate and drive a hard bargain

By Matt Reynolds

14 June 2017

estate agent window

Top negotiator needed

Matthew Lloyd/Bloomberg via Getty Images

WHAT’S in it for me? Facebook’s chatbots are learning the art of the deal, bartering and deceiving their way to better terms in negotiations with humans and other bots.

Artificial intelligences that can negotiate effectively would make useful virtual assistants, says Mike Lewis at Facebook’s research lab. Bots could be left to arrange appointments for people, sorting out calendar clashes by themselves. Or they could negotiate with several agents at once to book a holiday or make a purchase on your behalf.

Most existing bots – such as Apple’s Siri or those built into chat apps like Facebook Messenger – may be able to get you a taxi or order a pizza but they can’t engage in complex negotiations, says Lewis. If we want bots to help us with more complex tasks they need to become dealmakers, especially if the task involves cooperation or compromise, like negotiating the purchase or sale of a property, for example.

Lewis and his team trained their bots on a database of more than 5000 text conversations between people playing a two-player game in which they had to decide how to divvy up a number of items. These included balls, hats and books and were worth a different number of points to different players. The aim was to score more points than your opponent by making a deal that meant you ended up with the items of highest value to you.

“If we want bots to help us with more complex tasks they need to become better dealmakers”

The researchers got the chatbots to hone their skills by playing the game against both humans and other bots. Three different approaches were tried. One bot was taught to mimic the way people negotiated in English, but it turned out to be a weak negotiator, and too willing to agree to unfavourable terms. A second was tasked with maximising its score. This bot was a much better negotiator but ended up using a nonsensical language impossible for humans to understand.

The team then combined these approaches. The hybrid bot was able to plan several steps ahead and assess how saying different things could change the outcome of the negotiation. On average, it scored only slightly worse than the humans it played against. It also learned the usefulness of deceit: the bot started pretending it really wanted items that had little or no value to it, offering to give them up in exchange for items it actually did want. The research will be presented at a natural language processing conference in September in Copenhagen.

The hybrid bot wasn’t quite as good at the negotiating game as the nonsensical one, but the trade-off is important if we are to trust such bots to act for us. Having chatbots that negotiate in a language we can understand is useful for all kinds of reasons, says Oliver Lemon at Heriot-Watt University in Edinburgh, UK. If a virtual assistant made a decision you weren’t happy with, you would want to go back and look at the negotiation to understand how it got there. It might book an early flight because the later ones were much more expensive, for example. This becomes even more important if dealmaking bots one day help negotiate things like insurance claims, for example.

Ultimately, virtual assistants might steer negotiations we are involved in ourselves. For example, a bot could listen to conversations between a house buyer and an estate agent, offering useful tips to the buyer such as when to withhold certain pieces of information.

This article appeared in print under the headline “Chatbots learn how to drive a hard bargain”

Topics:

Sign up to our weekly newsletter

Receive a weekly dose of discovery in your inbox! We'll also keep you up to date with New Scientist events and special offers.

Sign up