Policy Optimization for Pedestrian Traffic Management by Surrogation of Simulation Models

Abstract

This paper proposes a method to swiftly develop crowd control policies that enhance the smooth flow of pedestrian traffic. The method employs a novel, gradient-based black-box optimization technique, which uses a neural network to create a surrogate model replicating a multi-agent simulator. This approach facilitates the rapid generation of improved policies by leveraging gradient information, thereby significantly decreasing the time required to determine optimal strategies. By treating the simulator and its evaluation function as differentiable entities, the method allows for quick policy adjustments guided by gradient information. The proposed method quickly and accurately obtains policy parameters, adapting seamlessly to different optimization strategies in various scenarios. Our methodology was tested by applying it to actual pedestrian flow data from Koshien Stadium in Hyogo, Japan. The real-world application not only confirmed the practical utility of our optimization technique in effectively managing crowd scenarios but also marked a significant advancement in public safety measures for densely populated areas.