In order to study the formation of symbol communication systems, we conducted an experiment based on a game introduced by Galantucci (2005). This involved pairwise communication in which the pairs engaged in a coordination task through an exchange of messages composed of a small set of geometric figures. We analyzed the transfer entropy—which is a measure of the information flow between two information sources or stochastic processes—between the figures used and the actions by individuals. Consequently, we confirmed that the transfer entropy reduced significantly across individuals, but not within each individual. Moreover, it correlated negatively with the performance of the task. From these results, we suggest that the transfer entropy appropriately shows the degree of formation of symbol communication systems. Then, we report the causal relationship between the uncertainty and the performances using a computational model based on reinforcement learning.