Synthetic Pedestrian Routes Generation: Exploring Mobility Behavior of Citizens through Multi-Agent Reinforcement Learning

Abstract

The data-driven city is improving the quality of everyday life for citizens and the efficient use of resources. While data-driven cities are benefiting from artificial intelligence technologies, one of the major challenges are the security, privacy, managing issues and costs. Furthermore, the training data for machine learning methods are in many cases not sufficiently available. Simulation tools are available for the designers or administration, and synthetic data generation based on real data is a promising approach on the development of new methods. This paper reports an exploratory research, targeting data generation with agent-based simulations combined with reinforcement learning. The study is still in its starting phase, but the conceptual outline and methodology has been created. The goal of this study is to establish a model to retrieve insights into the integration from a restricted data set from Google Maps and agent-based simulations. One of the primary observations is that intelligent agents, trained/reinforced paths instead of random behavior, show vastly different behavior than randomly acting agents. Evaluate the behavior according to empirical data, mostly because of the advantage of this method for complex behavior, where it would impractical to iterate through all possible paths, resulting in methodological issues. The paper exposes the following methodical challenges. (1) Realistic data sets will be large and full of decisions that appear to be random. To verify the collected, empirical data might lead to unreliable results, as actors will, as in real life, make different decisions according to circumstance. (2) The more complex a model has to be to be equivalent to the synthetic model, the more powerful the computational resources have to be. This leads to a dilemma between the model's complexity to make results comparable, and the simplicity to make computation realistic. Outline of the papers framework: (1) The generated data set challenges explored with the goal of providing realistic synthetic data. (2) A method for integration of synthetic and real-world data based on the reinforcement learning is developed. The premise is that implementing a reinforcement learning framework on top of multi-agent systems makes it possible to understand the mobility behavior of citizens. The paper explains the feasibility of both points, showing that this approach forms a solid basis for further investigation of the synthetic data generation for recognizing the mobility behavior of the citizens and enables the researchers to investigate the usage of reinforcement learning approach on the human mobility routes.