This paper describes learning in cheap talk games by computer simulations. Kawagoe and Takizawa made a number of experiments on certain cheap talk games. However, as is common to all experiments, subjects play each game for a small number of times and their actions are observed (their strategies are not observable). In this paper the cheap talk games are repeatedly played millions of times by those agents whose behavioral rules are apparent. The simulations where agents follow simple reinforcement learning rules reproduce the results of experiments in the long run, but it takes much longer period for the long-run equilibrium is realized. This suggests both the robustness of reinforcement learning and the fact that human subjects find long-run equilibria not by solo learning but with the help of inference. We believe our paper gives an illustration of how experiments and computer simulations can be used complementarily to understand human behavior in games.

Co-authors Atsushi Iwasaki and Kanji Ueda