情報処理学会 第81回全国大会講演要旨

Pedestrian behavior simulation is a difficult task to perform due to the performance requirements and the necessary information to learn meaningful behavior patterns. There are several instances when only having expert trajectories is not enough to obtain the expected knowledge. For example, under certain conditions, to traverse certain areas of unknown layout to reach concrete goals is difficult for trained agents, whilst for humans such information should be trivial to deduct. In order to avoid such situations, we first devised a method to improve the available graph information contained in the trajectory database. Then, we applied it to a system that generates agents capable of learning different behavior patterns and obtained promising results.