An Intentional AI for Hanabi
Published in Proceedings of the AAAI Conference on Artificial Intelligence, 2017
This paper addresses the challenge of creating AI agents that can effectively cooperate with human players in partially observable cooperative games. Using the award-winning card game Hanabi as a testbed, we developed an AI agent based on H.P. Grice’s maxims of communication, particularly the maxims of relation (relevance) and manner (clarity). Unlike previous AI approaches that optimize for AI-AI play, our intentional agent is specifically designed to communicate its intentions clearly to human partners through its hint-giving behavior. The research involved a large-scale experiment with 224 human participants playing 1,211 games, demonstrating that our intentional AI achieved statistically significantly higher scores when paired with human players compared to baseline approaches. The study provides empirical evidence that humans prefer and perform better with AI agents that exhibit intentional, goal-directed behavior and follow natural communication principles, contributing to the broader understanding of effective human-AI interaction in cooperative settings.
Recommended citation: Eger, M., Martens, C., & Alfaro Córdoba, M. (2017). "An Intentional AI for Hanabi." Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment.
Download Paper