Lazaridou., et al 2017 proposed a framework for language learning that relies on multi-agent communication. The agents in the framework were setup in a referential game where they communicated about many images. In this paper, we propose an experiment where agents develop a private language for referring to specified sentences given a set of sentences. The challenge is for the agents to learn a method of distinguishing differences between sentences and to develop a shared language to be able to refer to particular sentences by those distinguishing features. We will evaluate the agents' ability to accurately identify and differentiate the sentences. In addition, we will identify patterns in the methods that the agents develop to refer to the different types of sentences.
Keywords: Reinforcement learning, multi-agent coordination
New content is added all the time. Follow us on twitter for the highlights!