Learning efficient communication through reinforcement learning
This project concerns investigating how agents can learn to communicate efficiently for a given task using reinforcement learning, and touches upon cross-disciplinary aspects between machine learning, cognitive science and linguistics. As a suggested starting point we take the paper by McCarthy et al. (2021), which describes how humans invent and learn to use concepts for collaborating on a building task, where one gives instructions (the architecht) and one acts as a builder. We want to investigate if we can model similar emergent high-level concepts as humans invent, in our reinforcement learning agents.
Reference: Learning to communicate about shared procedural abstractions. William P. McCarthy et al (2021) https://arxiv.org/pdf/2107.00077.pdf