Recursive Metropolis-Hastings Naming Game

Generalization of Multi-agent Symbolic Emergence based on Probabilistic Generative Models

Abstract

A system in which a group of agents constructs a symbol system for communication and cognition in a bottom-up manner is called a symbol emergence system. A previous study extended a computational model of a naming game based on the Metropolis-Hastings naming game to a model involving three agents. In this study, we propose a model of the N-agent version of the naming game. The naming game is based on the Metropolis-Hastings naming game, based on the idea that the naming game among N-1 agents can be regarded as a sampling procedure of the Markov chain Monte Carlo method. In each iteration of the naming game, we computed the adjusted Rand index (ARI) of the category in each agent and the kappa coefficient of signs for each agent, and tested whether the behavior of each score changed using both synthetic and real image datasets. Each of the four agents successfully formed categories and shared signs.