Learning to Communicate using Deep Reinforcement Learning

The AI agents Alice and Bob walk into a bar. Alice wants Bob to order that new fancy drink everybody is talking about. Sadly, she has no word for the drink since it didn’t exist when her LSTM network was trained. She gets a pint. Though that story worked out just fine, it points to one of the many problems with the current paradigm of learning language from canned text. In an alternative paradigm, that is currently gaining traction in the research community, agents instead learn to communicate by solving tasks, e.g. ordering the correct drink for their buddy and getting reinforced if it succeeds.This leaves the agents with a functional language grounded in real objects and the ability to form new words when necessary. Further, the properties of the emerged artificial language can be used to answer important questions regarding our own language development.