Description: Humans are social animals living in physical environments. We not only pursue our personal goals in the physical world, but also routinely interact with one another. In addition, we can also spontaneously and robustly recognize the social goals (e.g., helping and hindering) and social relationships (e.g., friendly, adversarial) of others from observing their behaviors. There have been computational models for inferring agents’ social goals [1,2], but they assume that a fixed goal space is specified a priori. In this project, we will explore how to build a computational model that can discover the goals of social agents from observing their behaviors.
Specifically, we will focus on simulated social interactions in Heider-Simmel-type animations (see the figure below), where agents have different goals and relationships, and are interacting with the physical environment and with each other.
Code based on "Multi-Agent Adversarial Inverse Reinforcement Learning"
Lantao Yu, Jiaming Song, Stefano Ermon
- For code implementing MA-AIRL, please visit
multi-agent-irl
folder. - For the OpenAI particle environment code, please visit
multi-agent-particle-envs
folder.
- Go to branch
social_agents