[2O1-GS-8] Robot and real worlds

[2O1-GS-8-01] Identification of Multiple Intentions Implied in Human Flow by introducing EM to Inverse Reinforcement Learning

EMアルゴリズムと逆強化学習によるアプローチ

〇Masaharu Saito¹, Satiyo Arai¹ (1. Chiba University)

Keywords:Inverse Reinforcement Learning, Intentional Estimation

In recent years, human flow data has attracted attention for its application to disaster prevention, urban development, and safety in automated driving based on the analysis of human behavior. In the past, human flow data was generally used to understand the dynamic characteristics of groups of people, such as when, where, and how many people are present. On the other hand, Inverse Reinforcement Learning (IRL) is an approach to behavior analysis focusing on individual behaviors and their intentions. The estimated intentions are expected to guide evacuation, detect abnormal behavior, and elucidate the ecological behavior of organisms. However, conventional IRL assumes that all trajectories constituting human flow are generated by "a common intention. In reality, however, it is natural to assume that they contain multiple intentions.

Therefore, in this paper, we propose a method that enables the estimation of multiple intentions contained in human flow. The proposed method is a method (EM-MaxEntIRL) that introduces the EM algorithm into Maximum Entropy IRL (MaxEntIRL), which is one of the existing inverse reinforcement learning methods. The effectiveness of the proposed method was verified by computer experiments using airport behavior trajectory data reflecting actual human intentions to identify multiple intentions and estimate each intention.