Partner Capability Estimation for Task-Agnostic Adaptation in Ad-Hoc Teamwork
The paper tests agents that infer a partner’s hidden abilities across tasks, without population pre-training.
The authors introduce CE-CM, a Bayesian method for estimating task-invariant capability vectors during ad-hoc teamwork. It uses simulation-based sampling to update beliefs online and plan under hidden partner capabilities. A CE-CM-Div variant handles human sub-optimality by comparing hypotheses against diverse planner rollouts. In simulations, CE-CM reduced infeasible action assignments and adapted over time; in a 225-trajectory offline human study, CE-CM-Div improved estimates over CE-CM. ArXiv · AI/CL/LG's note
The authors introduce CE-CM, a Bayesian method for estimating task-invariant capability vectors during ad-hoc teamwork. It uses simulation-based sampling to update beliefs online and plan under hidden partner capabilities. A CE-CM-Div variant handles human sub-optimality by comparing hypotheses against diverse planner rollouts. In simulations, CE-CM reduced infeasible action assignments and adapted over time; in a 225-trajectory offline human study, CE-CM-Div improved estimates over CE-CM. ArXiv · AI/CL/LG's note
score 4