Maximum Entropy Population-Based Training for Zero-Shot Human-AI Coordination
About
We study the problem of training a Reinforcement Learning (RL) agent that is collaborative with humans without using any human data. Although such agents can be obtained through self-play training, they can suffer significantly from distributional shift when paired with unencountered partners, such as humans. To mitigate this distributional shift, we propose Maximum Entropy Population-based training (MEP). In MEP, agents in the population are trained with our derived Population Entropy bonus to promote both pairwise diversity between agents and individual diversity of agents themselves, and a common best agent is trained by paring with agents in this diversified population via prioritized sampling. The prioritization is dynamically adjusted based on the training progress. We demonstrate the effectiveness of our method MEP, with comparison to Self-Play PPO (SP), Population-Based Training (PBT), Trajectory Diversity (TrajeDi), and Fictitious Co-Play (FCP) in the Overcooked game environment, with partners being human proxy models and real humans. A supplementary video showing experimental results is available at https://youtu.be/Xh-FKD0AAKE.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Cooperative Multi-Agent Coordination | Overcooked-AI Asymmetric Advantages | Mean Reward177.5 | 22 | |
| Zero-shot Coordination | Overcooked-AI Coordination Ring | Mean Team Reward152 | 18 | |
| Zero-shot Coordination | Overcooked-AI Cramped Room | Mean Team Reward168.8 | 18 | |
| Zero-shot Coordination | Overcooked-AI Counter Circuit | Mean Team Reward65 | 18 | |
| Zero-shot Coordination | Overcooked-AI Forced Coordination | Mean Team Reward35 | 18 | |
| Coordination | Overcooked Cramped Room layout v1 | SP203 | 14 | |
| Multi-agent coordination | Overcooked Coord. Ring | Average Return40.3 | 10 | |
| Human-Agent Coordination | Overcooked Counter Circuit (human evaluation) | Average Score76.19 | 9 | |
| Human-Agent Coordination | Overcooked Multi-strategy Counter (human evaluation) | Average Score64.02 | 9 | |
| Multi-agent coordination | Overcooked Asymm. Adv. | Return127.4 | 8 |