CE-CM: New AI method lets agents estimate hidden partner abilities in ad-hoc teams
AI learns to collaborate with unknown partners using just a few tasks, no pre-training needed.
Current ad-hoc teamwork systems assume known partner capabilities and fixed tasks. A new paper from Peter Tisnikar and colleagues introduces CE-CM (Capability Estimation via Contextual Models), an approximate Bayesian method that infers a partner's hidden capabilities from just a few tasks. By reframing collaboration as joint planning with decentralized execution, CE-CM uses simulation-based sampling to estimate task-invariant capability vectors and induces contextual multi-agent MDPs for planning. It requires no population pre-training and refines beliefs online.
The method’s extension, CE-CM-Div, accounts for human unpredictability by evaluating capability hypotheses against diverse planner rollouts rather than a single optimal trajectory. In simulated experiments, CE-CM rapidly recovered hidden capabilities, reduced infeasible action assignments, and adapted to changes over time. A human study of 225 trajectories from 15 participants showed CE-CM-Div substantially improved capability estimates over the baseline. This capability-based modeling promises interpretable, task-agnostic representations for robust human-AI teaming.
- CE-CM infers hidden partner capabilities using Bayesian simulation-based sampling without pre-training.
- The method adapts online from just a few tasks, reducing infeasible action assignments.
- CE-CM-Div accounts for behavioral diversity, improving estimates in a human study with 225 trajectories.
Why It Matters
Enables autonomous agents to collaborate effectively with unknown human partners, advancing practical human-AI teamwork.