Facebook AI to Open-Source “Negotiation Bots”: Are We Ready?
Facebook AI Research (or FAIR) just announced results of its work around “negotiation bots” and have also released the open-source code relating to their research. These bots are designed to carry out basic negotiation skills with humans and other bots.
This is a new frontier because negotiations are complex. It requires reasoning, balancing goals between multiple parties, and compromise. All very human requirements. According to the post,
“Negotiation is simultaneously a linguistic and a reasoning problem, in which an intent must be formulated and then verbally realized.”
Today’s chatbots can only perform simple tasks like ordering or booking from a restaurant. So this is seen as an important first step in achieving true human-like conversations with robots or chatbots.
Machines 1. Humans 0…?
As you would expect, FAIR chatbots although promising, their interactions and ability to negotiate is still limited. However, FAIR researchers expedited the process by crowd-sourcing negotiations across pairs of people. In order to move past just copying those negotiations, FAIR trained the bots to practice thousands of simulations against itself, reinforcing good outcomes with rewards.
As a result, when the chatbots were tested against actual human beings, most people did not realise they were speaking with a chatbot. In addition, they achieved the capability to converse fluently in English while negotiating for themselves.
You’ve got to learn how to crawl before you can walk
But don’t get excited just yet. The simulation test and negotiation was limited and narrow in scope: bargaining for basketballs, books and a cowboy hat on a point system.
The important takeaway from this experience is that AI and AI chatbots learn through trial & error. Which makes sense, to become human is to learn like one.
FAIR chatbots grew increasingly smarter and achieved its goals through FAIR’s supervised learning and reinforcement learning methodology. It was deliberate, measured and iterative. In other words, learning through doing. AI chatbots cannot be smart from Day 1. It takes time, patience, and an extraordinary understanding of the underlying challenge. Not just for humans, but also for…chatbots.