Facebook has recently published a brand new paper, that suggests that the company has been experimenting with teaching the art of negotiations to artificial intelligence. Interestingly, this publication comes on the heels of Facebook hiring Apple’s chief of natural language understanding.
Conversational AIs are going to be extremely important in the coming future. These AIs are capable of conversing with humans and as can be predicted, can be used to drive sales. However, the one major disadvantage they have as compared to a human is their inability to negotiate.
In order to achieve this, Facebook took the help of game theory, combining it with deep learning and implementing techniques used in AIs that play games to create machines that can potentially negotiate. Researchers started off with an imaginary scenario that had Humans on Amazon’s Mechanical Turk presented with an explicit value function. Next up, they were asked to negotiate using natural language and attempt to obtain the maximum reward from a pot. The person at the other end of the negotiating table — well, it wasn’t a person but a bot!
The negotiations could not exceed 10 rounds of dialog at the very most, so there was that constraint as well. And the machines came out pretty admirably, learning many classic tricks in the process including the placement of false emphasis on low value items. Later on, these items could be used to exert pressure on the person that was being bargained with.
The technique underlying this whole process is basically that of a decision tree. A decision tree is a tree which can denote all possible outcome after a possible step. Humans can be said to operate according to a decision tree as well but in our case, the possibilities are mind bogglingly large due to the element of unpredictability. For language for instance, we can state the same statement in a very, very large number of ways.
Facebook also modeled its agents using the conversations between two humans. Then, these bots were made to negotiate with each other. Finally, by reinforcement learning (giving rewards for good negotiation), the bots were trained in what is the correct method. The system isn’t perfect yet but there is a lot of scope in the field.
