ABSTRACT

This study investigates the mechanism by which individuals learn to associate signals with meanings in a way that is agreeable to everyone, and thereby, to collectively produce common and stable signaling systems. Previous studies suggest that simple learning algorithms based on local interactions, such as reinforcement learning, sufficiently give rise to signaling systems in decentralized populations. However, those algorithms often fail to achieve optimal signaling systems. Under what condition do suboptimal signaling systems emerge? To address this question, we propose a multi-agent model of signaling games with three parameters – memory length, the complexity of communication problems, and population size – as potential constraints imposed on the collective learning process. The results from numerical experiments suggest that finite memory leads to suboptimal signaling systems, characterized by redundant signal-meaning associations. This paper concludes with discussions on the theoretical implications of the findings and the directions of future research.

摘要

这项研究调查了个人学习以每个人都同意的方式将信号与意义联系起来的机制，从而共同产生共同和稳定的信号系统。先前的研究表明，基于局部交互的简单学习算法，如强化学习，足以在分散的人群中产生信号系统。然而，这些算法通常无法实现最佳信令系统。在什么条件下会出现次优信号系统？为了解决这个问题，我们提出了一个信号游戏的多代理模型，它具有三个参数——记忆长度、通信问题的复杂性和人口规模——作为对集体学习过程施加的潜在约束。数值实验的结果表明，有限的记忆会导致次优的信号系统，其特征是冗余的信号意义关联。本文最后讨论了研究结果的理论意义和未来研究的方向。