Creation of meta-model for agent-based simulation using machine learning approach

Abstract

The ability to predict and model crowd behavior is an overwhelming and challenging task. This paper investigates the viability of neural networks in their ability to learn crowd patterns from video data. We tackle the problem of simulating crowd behavior by using neural networks to tackle the problem and automatically understand the physics and patterns behind the data. Neural network’s ability to generalize through training data is one that we also aim to leverage in our simulation. We design a simple fully-connected neural network with one hidden layer and tanh activation function. To evaluate its performance, first the neural network learns appropriate weights and biases from a video dataset in Switzerland following which its performance is validated against video data from the United States. The neural network is able to outperform Social Force Model and the Shortest-Path model, achieving a lower score for mean density errors. However, the difference between the Social Force Model and the Neural Network is shown to not be statistically significant. Furthermore, it performs slightly worse with regards to velocity field errors when compared to the Social Force Model. Despite the mixed results, the results obtained demonstrate that a neural network could model crowd dynamics of another scenario in a different context if the crowd behavior patterns are similar. Further investigation, with larger and more varied datasets and different neural network architectures are needed to showcase their capabilities in the field.