Facebook AI researchers are developing techniques to teach bots how to carry out negotiations with humans and others bots. Results of initial work in this area will be open-sourced today, and initial results will be submitted to Arxiv for peer review.
To train negotiation bots, FAIR used supervised learning, first observing people carry out negotiations, then training neural networks to imitate people. The bots were tested in online conversations with people and bots, and researchers found that the result was as good or better than people.
Researchers call the technique for bots to carry out negotiations “dialog rollouts.” Negotiation requires a certain amount of thinking ahead, and dialog rollouts use a dialog model to simulate future conversations.
Each negotiation was carried out by assigning numeric value to items and having two bots split a handful of items, like balls, books, and bats. Quitting wasn’t an option, because quitting would mean both bots earn zero points.
The news today is FAIR’s latest effort to share insights with a larger community of businesses and developers interested in bots and artificial intelligence. To promote different forms of dialogue and grow the use of bots that can hold conversation, last month FAIR launched ParlAI.