Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Learning to Collaborate with Unknown Agents in the Absence of Reward
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes STUN, which infers latent rewards of unknown agents via kernel-density Bayesian inverse learning (KD-BIL) and adapts goal-conditioned MARL policies to them without reward access or online retraining.
Problem
Training AI agents to collaborate with unknown agents whose reward is unavailable to the AI agents or undefined by the unknown agents themselves, where standard MARL and ad hoc teamwork require a known shared reward.
Approach
Use kernel density Bayesian inverse learning to estimate the posterior over the unknown agents' latent reward parameters from observed trajectories without refitting policies, feed the estimate into a family of goal-conditioned policies pre-trained over sampled surrogate rewards, and show that Q-learning with unbiased inferred rewards converges to optimum.
Key results
- On the 3s vs 5z map, STUN achieves a normalized teaming score of 98.9 (reported as 99.2 in text) versus 65.6-88.7 for baselines across 8 unknown agents.
- On super-hard maps such as 27m vs 30m and MMM2, STUN improves unknown-agent reward by up to 50%.
- KD-BIL posterior estimates correlate with ground-truth latent reward parameters on 3s vs 5z and corridor maps; STUN adapts within 5-10 epochs to non-stationary unknown agents.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use STUN instead of the strongest baseline, FBA-A | Collaborative agents teamed with unknown agents · SMAC 3s vs 5z | average reward: 2.07 -> 2.31 | Helped |
| Remove adaptive policy update and perform online reinforcement learning using the inferred reward (ablation) | STUN · MPE Predator-Prey | teaming performance | Hurt |
| Remove active goal inference and condition policies on fixed reward parameters (ablation) | STUN · MPE Predator-Prey | teaming performance | Hurt |
| Use STUN instead of baselines on super-hard maps | Collaborative agents teamed with unknown agents · SMAC super-hard maps | teaming performance: up to 50% | Helped |
Surprising findings
- The paper argues that using the maximum a posteriori estimate of the latent reward does not guarantee optimal Bellman convergence, whereas an unbiased posterior estimate is necessary.
Limitations
- Evaluated only on redesigned MPE and SMAC environments with built-in or surrogate agents.
- Assumes fully observed or specific observation settings; restricted observations and non-stationarity are stated as future work.
- Convergence guarantees rely on finite MDP assumptions for the Q-learning analysis.
Experimental setup
Datasets: MPE Predator-Prey (redesigned) SMAC (redesigned)
Benchmarks: redesigned SMAC maps (3s vs 5z, 5m vs 6m, 6h vs 8z, 27m vs 30m, corridor, MMM2) redesigned MPE Predator-Prey
Metrics: average latent reward normalized teaming score
Baselines: MAPPO IPPO COMA MAA2C IA2C fixed-behavior agents (FBA-C, FBA-B, FBA-A) multi-task learning agents
Builds on: Bayesian inverse reinforcement learning KD-BIL MAPPO dec-POMDP
Closest papers by meaning
- Multi-agent cooperation through learning-aware policy gradients (ICLR 2025 · similarity 0.87)
- Cautiously-Optimistic Knowledge Sharing for Cooperative Multi-Agent Reinforcement Learning (AAAI 2024 · similarity 0.87)
- NestRL: A Nested Training Regime for Mutual Adaptation in Human–AI Teaming (NeurIPS 2026 · similarity 0.87)
- Training Generalizable Collaborative Agents via Strategic Risk Aversion (NeurIPS 2026 · similarity 0.86)
- Adaptively Coordinating with Novel Partners via Learned Latent Strategies (NeurIPS 2025 · similarity 0.86)
- N-agent Ad Hoc Teamwork (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.