Emergence of compositional sequential codes among RL agents

EE376A (Winter 2019)

Govardana Sachithanadam Ramachandran


Communication in living things arise out of necessity. Animals associate each sound they make to an intent, for example dogs bark could mean one of few intents such as warning, intimidation, mating call etc.,. But human languages are compositional, with fewer than 5000 words one could describe anything from cardiac surgery to Hamlet of Shakespeare. Compositional rich language have high information capacity and generalize to unseen scenarios, an example would be kids making up new words by morphologically associating known words. This project analyzes and find ways to encourage the emergence of these compositional language –which be can interpreted as source coding of intents– in a simulated multi AI agent Reinforcement Learning population. In this work we process optimization schemes to train a neural network for discrete sequential code communication among RL agents

Full report can be found here


It was quite challenging and gratifying to explaine my project to the students of Nixon Elementary School. I was faced with the challenge of explaining concepts from Information theory and reinforcement learning. I tried to distill the core concepts with ample use of visual aid, cartoon characters and videos, which the kids could relate to .
One example, is explaining the fact that human language have higher information capacity compared to their animal counterpart. For which I played various sounds a fox makes and their corresponding intent

For the compositionality of human language, I encouraged them to come up new sentence by rearranging words, with below sentences as example

Another example, is explaining the concept of training a reinforcement learning model. For this I gave the analogy of training a dog to do tricks, explaining the concept of reinforcing behavior with rewards in the form of treats

Overall it was fun. One key take away is that, it was easier to get the concept to across if I engaged their parents as part of the presentation, as they were able to abstract the concept even further so that the kids could understand