go back

Measuring cooperation in Hanabi agents using information theory

Christiane Wiebel, Matti Krüger, Patricia Wollstadt, "Measuring cooperation in Hanabi agents using information theory", International Conference on Hybrid Human-Artificial Intelligence (HHAI 2023) , 2023.


We propose to use a novel measure from information theory, called synergistic information, to measure cooperativity between two interacting agents. The synergy quantifies joint effects of two inputs towards a target that can not be obtained from either input alone. We argue that this definition of the synergy corresponds to popular notions of cooperation from prior art. In particular, we propose to estimate the synergy between two agents' actions with respect to the current task's goal. We here apply this measure to rule-based agents playing the popular cooperation benchmark, Hanabi. We estimate the synergy between agents' actions and whether a point was scored in a round. We estimate the synergy from agents with different cooperative abilities and show that the synergy correlates with the total number of points scored in a game. We conclude that the synergy is a promising measure of cooperativity for interacting agents with potential applications also in human-machine interaction.

Download Bibtex file Download PDF