Facebook Is Teaching AI Chat Bots How To Negotiate With Real Humans
Facebook's chatbots, or artificial intelligence-powered assistants, are slowly gaining more human-like traits.
Facebook's chatbots already help people order pizzas, accept payments, and even ask for translations, but while they're already pretty useful, Facebook know it can do more with the help of its developments in AI.
Enter a bot that can negotiate, strategize, and plan ahead in a conversation, which Facebook announced on Wednesday, June 14. Teaching bots dynamic conversational abilities akin to the way humans talk to each other has left scientists, engineers, and software developers mostly stumped.
Why? Well, because the task presents pretty steep hurdles. For robots to talk with and understand humans in a seemingly natural way, they need to be equipped with a broad lexicon, an advanced language capacity, and accurate adaptability, for good measure. But researchers at the Facebook Artificial Intelligence Research, or FAIR, have developed a unique way for bots to negotiate, a complex conversational skill even some humans sometimes fail at.
Facebook Has Managed To Develop Bots The Can Negotiate
Thus far, Facebook's bots have managed to perform relatively menial tasks, with the smart assistants able to hold short and quick conversations with users who need to get something done. But bots that can talk to real humans in a meaningful way proves difficult to achieve, which Facebook admits.
"[It's] challenging because it requires a bot to combine its understanding of the conversation with its knowledge of the world, and then produce a new sentence that helps it achieve its goals," the company wrote in a blog post.
To make a bot that can negotiate, researchers at FAIR looked at actual negotiation conversations between two people. These people had to make a deal over a set of items, and each person had different sets of values for each. Then the individuals had to negotiate how to split the objects between them. Facebook then trained a recurrent neural network to negotiate based on the actions of those human subjects.
But FAIR went a punch above imitation — it also trained the system to learn how to achieve negotiation goals, reinforcing good outcomes when they happened. Then Facebook tested the negotiating bots on real people, most of whom Facebook claims didn't realize they were interacting with a bot, thereby proving the bots' conversational skill too.
To equip the bots with that level of language command, Facebook employed so-called "dialogue rollouts," which allows for long-time dialogue planning. Basically, the bot predicts how a conversation can potentially go.
Facebook Bots: Learning To Bluff
Over time, Facebook claims that the bots even learned how to bluff, pretending to care for a result they didn't actually want just to have an upper hand later in the negotiation.
"This behavior was not programmed by the researchers but was discovered by the bot as a method for trying to achieve its goals," says Facebook's blog post.
Possible Implementations Of Facebook's Negotiating Bots
Of course, deciding how to split up a set of objects doesn't sound like much of a big deal, but for bots to actually decide and attempt to reach the best outcome possible, perhaps even through bluffing, still stands as an impressive achievement. Facebook hopes these negotiating bots could help real-world conversation going forward, possibly having the bot negotiate something as simple as meeting times or something as complex as conducting a sale or a business deal.
Facebook says that these bots stand as progress in its efforts to build a personalized digital assistants. The company is also publishing their research on the work, in addition to releasing open-sourced code.