Recently, the importance of the walkability in the urban areas is re-evaluated., and many kinds of pedestrian behavior models have been proposed. However, most of them do not deal with the multi-modal situation. Usually, the pedestrians and the vehicles on a street recognize each other, and the behavior of the one side affects the other side. When this effect is formulated by the recursive logit model, the estimation needs very expensive iterative calculation. To tackle this problem, this research proposed a new method based on adversarial inverse reinforcement learning. This model learns the link value function for each mode in a single training process, which much reduces the computational cost. In the case study in Matsuyama, the same level of estimation as RL can be done in much smaller time.