Facebook is working on to teach Machines the art “Negotiation with Humans”

The brand new paper published recently by Facebook is connoting the experimentation with teaching the art of negotiations to company’s AI (artificial intelligence). Even though it is not considered as important as other big companies like Apple, Amazon and Microsoft, Facebook’s papers have also stimulated interest in dialog system. It may seem easy but the dialog is very hard which can be explicitly seen in the contemporaries namely Siri, Alexa and Cortana, it not only requires speech recognition but much more efforts in order to provide a remarkable experience to users.

Facebook’s Subtle Art- Negotiation with Humans:

It is very evident that conversational AI’s are here to stay in future and can be used to drive the sales which are why they hold extreme importance. The one disadvantage they carry is their inability to negotiate which becomes a major hindrance to overcome. And Facebook with their deep learning and techniques, taking help from the game theory, has created machines that can negotiate with humans potentially. They have created machines capable of complex bargaining by using the rollout techniques which are more commonly used in game playing AI’s.

Facebook researchers started by dreaming of an imaginary negotiation scenario having humans on Mechanical Turk of Amazon with a clear value function. These humans were asked to negotiate in natural language with an objective of maximizing the reward from a pot. The one thing to be taken into consideration was not to exceed the limit of rounds of dialog which was set as 10 as nobody would receive a reward if this constraint was met.

Both the agents had different preferences and therefore, they had to develop a dialog so as to sort out the objects and decide which objects should be given to which agent considering their preferences. In this process of interaction, machines naturally learnt many classic negotiation tricks one of which is placement of false emphasis on low value items.

This technique basically takes the form of a decision tree which is already a crucial component of several AIs. To define it, a decision tree is a tree which is able to denote all outcomes which can possibly occur after a possible step. These decision trees allow us to take decisions by modeling future states from the present. Albeit humans are also said to be operating according to a decision tree but with us, the possibilities are extremely large in number because of the element of unpredictability.

It can also be understood with the help of the game Tic-Tac-Toe, in the game we have finite options i.e. places where X can be placed, at any given point. So, each turn has a possible value. But dialog is much more complex than such games as the former does not advance from a finite number of outcomes i.e. there exist infinite human responses for any question.

As it is difficult to make the machines interact with humans, Facebook practised its models on negations of humans with each other. Once this happened, pairs of machines were also set to negotiate with the help of reinforcement learning where they were provided with rewards at the end of every good negotiation and that is how the bots were trained.

This system cannot be stated perfectly yet but carries great vitality in the field.

Read the full paper here