Deep Reinforcement Learning for Pedestrian Guidance

Abstract

In large-scale events where many people gather, providing them with appropriate, efficient, and safe guidance about where to proceed is critical to ease congestion. We can evaluate guidance candidates using a pedestrian flow simulator to find appropriate guidance. However, evaluating many candidates by simulation requires high computational cost, which prohibits real-time guidance. We propose a method that finds appropriate guidance in real-time for observed situations based on deep reinforcement learning. Our proposed method learns a function that outputs appropriate guidance given the observed situation to minimize the average travel time of pedestrians. The difficulty here is that the real-world measurements of pedestrian travel time are limited due to privacy issues since it tracks individuals. Though our method uses only the observation obtained without locating specific individuals: the number of pedestrians who are moving on roads, it is guaranteed by Little’s law to be equivalent to minimizing the average travel time. Our experimental results for unknown pedestrian flow show that our proposed method outperforms rule-based controls, and its guidance is as effective as one selected from many candidates by repeated simulations with massive computational cost.