Agents with Theory of Mind for
Intelligent Collaboration (ATOMIC)
DARPA Artificial Social Intelligence for Successful Teams (ASIST)
Internship @ SRI International
| Objective
We are interested in understanding how people interact with each other in collaborative settings, especially in situations in first encounters and where individuals know little about their teammates or adapt the goals of others as perceiving others' behavior. We approach this problem using Multiagent Inverse Reinforcement Learning via Theory of Mind (MIRL-ToM), where the goal is to infer the reward functions guiding the behavior of each individual given the team trajectories during task performance.
| Two-Phase Approach
For each agent, we first use ToM reasoning to estimate a posterior distribution over baseline reward profiles given their demonstrated behavior. We then perform MIRL via decentralized equilibrium by employing single-agent Maximum Entropy IRL to infer a reward function for each agent, where we simulate the behavior of other teammates according to the time-varying distribution over profiles.
| Search-and-Rescue Operation
We evaluate our approach in a simulated 2-player search-and-rescue operation (with PsychSim platform) where the goal of the agents, playing different roles (Medic and Explorer), is to search for and evacuate victims in the environment.
Our results show that the choice of baseline profiles is paramount to the recovery of the ground-truth rewards. We also show that the MIRL-ToM approach is able to recover the rewards used by agents interacting both with known and unknown teammates.
The work was submitted to AAMAS'23 and is currently under review.
Internship Advisor: Dr. Pedro Sequeira
H. Wu, P. Sequeira, and D. V. Pynadath, "Multiagent Inverse Reinforcement Learning via Theory of Mind," The 22nd International Conference on Autonomous Agents and Multiagent Systems (AAMAS'23), 2023, London, UK. [Full Text] [Preprint Version]