Facebook’s AI Started Speaking Its Own Language And Negotiating Like Humans


Short Bytes: Researchers at Facebook trained their AI bot which they call ‘dialog agent’ to negotiate like humans and even converse in natural English language to the extent that even people find it hard to know if they are talking to an AI. They have open sourced the dataset and code of their machine learning algorithm.

The AI chatbots can do a lot of things for us. They can tell us a joke, book a table, or take slangs we throw at them as if they’re unaware of it. But the researchers at FAIR (Facebook Artificial Intelligence Research) think the AI can do far more than just ordering groceries; they can create their own language and even negotiate with people.

New research published by FAIR describes how the bots, referred as ‘dialog agents’ were trained to develop the ability to negotiate. Through the training process which involved machine learning algorithms and intelligence strategies, agents were able to bargain against each other for the things they liked more.

Also Read: Do I Really Own My Facebook Photos? Will Facebook Sell Them?

The agents developed the ability to negotiate for a long time without giving up, use tactics which weren’t programmed by the researchers, and produce more natural language.

facebook ai negotiation
Image: Facebook Research

Much to the surprise of the researchers, the dialog agents started negotiating in an entirely new language when they were left to converse without human attention. However, the researchers prevented this situation, making one of agent give responses in the English language as a human would do.

The supervised pre-training involved feeding an agent with a set of negotiations crowdsourced from people, and agents were able to imitate humans in situations. Later, the researchers used reinforcement learning so that the agents could learn thousands of negotiations on its own and get rewarded.

Supervised learning was used to teach how to map between language and meaning. On the other hand, reinforcement learning helped the agent with what to say and when.

When placed in front of humans (online) who didn’t know who they were chatting with, an agent matched negotiation skills of humans along with English as fluent that the people were left unsuspecting about its existence.

The new algorithm can help developers create advanced bots that reason, converse, and negotiate better, the researchers said in their blog post. These are all vitals to making a personalized digital assistant, and maybe, we could even see AI brokers getting cheap home deals for us.

You can find more details about the research titled ‘Deal or no deal? Training AI to negotiate’ using this linkIts open source code is now available on Github.

Got something to add? Drop your thoughts and feedback.

Also Read: China’s ‘Unhackable’ Quantum Satellite Sends Its First Successful Message
Aditya Tiwari

Aditya Tiwari

Aditya likes to cover topics related to Microsoft, Windows 10, Apple Watch, and interesting gadgets. But when he is not working, you can find him binge-watching random videos on YouTube (after he has wasted an hour on Netflix trying to find a good show). Reach out at [email protected]
More From Fossbytes

Latest On Fossbytes

Find your dream job