Leveraging Cognitive Bias for Zero-Shot Human-AI Coordination

Shray Bansal, Jin Xu, Miguel Morales, Jonathan Streater, Ayanna Howard, Charles Lee Isbell

Keywords: zero-shot human AI coordination, reinforcement learning, cognitive bias

TL;DR: Include cognitive biases in reinforcement learning to develop agents capable of zero-shot coordination with humans.

Abstract:

Recent advances 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 of optimal decision-making and behavior homogeneity when interacting with humans. Although we can train models to learn the best responses to human behavior given a large corpus of human-human interaction, the cost of collecting this data can be prohibitive. We show how even without such data, we can use our knowledge of biases and limitations in humans to construct a technique that can coordinate with humans. To do this, we present
an approach that learns models partnered with reinforcement learning agents that incorporate human behavioral biases. We evaluate this method in the fully-cooperative game Overcooked. Our results show an improvement when incorporating this bias with methods that do not include this bias in their agent population.

Submission Number: 16