Multi-agent Reinforcement Learning

Zero-shot Human-AI coordination

Recent advances in multiagent reinforcement learning have enabled artificial agents to work together effectively in complex environments. However, these agents often face difficulties when collaborating with humans. This is partly because they make assumptions about how humans make decisions and behave, which may not always be accurate.

Our research focuses on addressing the challenge of coordinating between humans and AI without prior knowledge of human behavior in a given task. We explore the problem of zero-shot human-AI coordination, where an agent is paired with a human partner in a cooperative task without access to data on human behavior. We explore using known cognitive and behavioral biases (e.g. Confirmation Bias, Anchoring Bias, Loss Aversion) to generate a set of agents to help enable the best response agent to coordinate with humans.

We employ the Overcooked environment developed by Carroll et al.. This environment offers a mix of strategy and motion coordination challenges, making it particularly suitable for training deep reinforcement learning algorithms.

For more information, please check out our recent publications (Bansal et al., 2022; Bansal et al., 2024).

Zero-shot human-AI coordination problem in the Overcooked environment.

References

2024

  1. Reinforcement Learning with Cognitive Bias for Human-AI Ad Hoc Teamwork
    Shray Bansal , Jin Xu , Miguel Morales , and 3 more authors
    In Coordination and Cooperation for Multi-Agent Reinforcement Learning Methods Workshop , 2024

2022

  1. Nash Equilibria in Bayesian Games for Coordinating with Imperfect Humans
    Shray Bansal , Miguel Morales , Jin Xu , and 2 more authors
    In Workshop on Strategic multi-agent interactions: game theory for robot learning and decision making at CoRL , 2022