Theory of Mind Modelling in Symmetric Multi-Agent Environments
Description: Theory of Mind is the ability that humans have to infer others' mental states, such as beliefs, desires and intentions. Early work in machine theory of mind mainly attempt to design agents that model mental states of others as passive observers [1]. In contrast, Sclar et al. [2] propose to model machine theory of mind in a more general symmetric scenario. In SymmToM, agents move in a gridworld and communicate to gain all available information. Since hearing is limited to its neighbor cells, they must guess what happened beyond this range. Performing this task effectively requires theory of mind. The goal of this project is to and explore ways of modelling theory of mind in SymmToM.
Goal:
- implement the reinforcement learning baselines for SymmToM in [2]
- explore and implement new ways to model theory of mind
- test different types of RL algorithms
- perform different types of analysis
Supervisor: Matteo Bortoletto
Distribution: 10% literature, 70% implementation, 20% analysis
Requirements: self motivation, strong programming skills in Python and PyTorch and/or Jax, knowledge of deep learning and reinforcement learning.
Literature: [1] Rabinowitz, Neil, et al. "Machine theory of mind." ICML 2018.
[2] Sclar, Melanie, Graham Neubig, and Yonatan Bisk. "Symmetric machine theory of mind." ICML 2022.
