情報処理学会 第81回全国大会講演要旨

We present a model for pedestrian behavior using a novel technique we called contextual action multiple policy inverse reinforcement learning (CAMP-IRL). Scripted simulations are often limited in their flexibility, so as a possible solution we can learn from data obtained in real scenarios using machine learning techniques and generate behavior patterns. Using this philosophy, we implemented a behavioral agent model that obtains different behavior patterns obtained training data, generating different trajectories depending of their goals and the environment. Our method also provides a way to switch dynamically between behaviors, thus being a robust way for agents to behave realistically.