Cognitive Bias for Human-AI ad hoc Teamwork

Advancements in multiagent reinforcement learning have enabled artificial agents to coordinate effectively in complex domains; however, these agents can struggle to coordinate with humans, in part due to their implicit but inaccurate assumptions about optimal decision-making and behavioral homogeneity while interacting with humans. Although we can train models to learn the best responses to human behavior using a large corpus of human-human interaction, the cost of collecting this data can be prohibitive. We demonstrate how, evenwithout such data, we can leverage our knowledge of biases and limitations in human behavior to develop a technique for effective human-agent coordination. To do this, we present an approach that trains an RL agent by best responding to a pool of other agents that incorporate human behavioral biases. We evaluate this method in the fully-cooperative game Overcooked. Our results show an improvement when incorporating these biases compared to methods that do not account for these biases within their agent population.