[1Win4] Poster session 1

[1Win4-27] A Computational Analysis Method for the Emergence of Communication with Temporally Extended Communication Games

〇Naoki Inoue¹, Kei Wakabayashi¹ (1.University of Tsukuba)

Keywords:Emergence of Communication, Artificial Intelligence

Understanding the mechanisms underlying the emergence of communication is a crucial issue across various fields, and computational analysis has been used to compensate for the lack of physical evidence. A common approach represents sender and receiver actions and payoffs in a matrix and analyzes them using game theory. However, most models assume single-shot interactions and do not account for environments where signaling and response effects propagate over time. In such cases, payoff matrices cannot be easily defined due to reward delays, requiring estimation methods that remain underexplored. This study defines "Temporally Extended Communication Games" to address this issue and proposes a method for analyzing its payoff structure. We create three policies by constraining reinforcement learning agents' signaling and response behaviors and estimate the environment’s payoff matrix from policy combinations. Experiments show that this representation evaluates policy optimality, though agents may fail to learn communication policies even when they are optimal.