Simulating whole city traffic with millions of multiple vehicle agents

Simulating large city traffic in microscopic is one of the fundamental technology to understand the nature of urban traffic. Also, nowadays, it is recognized that the variousness of driver behaviors plays an important role in overall traffic. In order to provide an experimental environment to simulate large city traffic with multiple driver behaviors, we designed and developed a large-scale microscopic traffic simulator that simulates traffic of millions of vehicles. We adopt multi-agent simulation framework and have developed the traffic simulator based on a Java-based massive agent-based simulation platform, ZASE. We have so designed the simulator that users can execute traffic simulation by varying driver models and by varying the mixture ratio of the driver models. We apply the traffic simulator to two road networks, a bottleneck road and a road network of the Kyoto city. Our agent-based traffic simulator is found to reproduce the fundamental characteristics of traffic flow, the q-ρ fundamental relation with metastable state. We carry out a simulation of the Kyoto-city road network of 32 thousand links and 22 thousand nodes with 0.89 millions of vehicles. In this paper, we clarify the design of the large-scale agent-based traffic simulator and discuss simulation results. Introduction Motivation Automobile traffic causes a wide range of traffic problems, such as traffic congestion problem, traffic safety problem, city’s air pollution, noise problems, and global warming issue. These traffic problems in metropolis are major problems in today’s modern society. Solving these traffic problems in simulation attracts more attention since simulation requires low cost and is free from risk comparing pilot programs. A significant amount of work has been done to model and to simulate traffic since the pioneering work by Lighthill-Whitham and Richards in 1950s (Lighthill & Whitham 1955; Richards 1956). In 1990s, studies in modeling traffic from the viewpoint of the many-body interacting system has been developed rapidly (see for example (Chowdhury, Santen, & Schandshneider 2000; Helbing 2001)). Inspired by these microscopic traffic models, today, lots of multi-agent traffic simulator which models huCopyright c © 2008, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. man driving behavior applying multi-agent simulation technologies (Dresner & Stone 2005; Yamashita et al. 2005; Cetin, Burri, & Nagel 2003; Cetin & Nagel 2002). A common issues in these prior works are two fold. Firstly these multi-agent simulators adopt some macroscopic models in calculating the vehicle speed such as queue model and cellular automaton model. Thus these simulators are not completely ”agent-based”. Secondly, the number of vehicles in a traffic system generated in these models are up to 100 thousands, which is insufficient to simulate traffic in metropolitan city such as New York, London, and Tokyo, where 2 to 4 million cars are owned. Recently Yamamoto et al. has developed a Javabased massive agent-based simulation platform named “ZASE” (Yamamoto, Tai, & Mizuta 2006). By the use of the thread pooling technology, ZASE enables to host over millions of agents on a single personal computer. Thus, ZASE provides a running environment for users to execute massive agent-based simulation such as auction simulation, mega-scale human navigation, and artificial market. Applying ZASE to large-scale traffic simulation is straightforward and thus it is expected that ZASE realize a large-scale agent-based traffic simulator. ZASE not only provides basic functions to host over millions of agents, but also provides capabilities for massive agent-based simulation applications to run on multiple computers connected with a high performance network. By using ZASE on a PC cluster, we expect that we can simulate further higher number of vehicles. In this paper, we propose a new traffic simulator for simulating the metropolitan-area scale traffic in microscopic, which we refer to as the ”Metropolitan-Area Scale Microscopic Traffic Simulator” (MASMITS). The key property of MASMITS is that it deals the largest-scale traffic with millions of driver agents, which has become possible powered by ZASE. Also MASMITS is the first traffic simulator to simulate all driver behavior with the granularity of vehicles. Summary of contributions This paper contributes to the literature from application side. Our contribution is to offer an experimental environment on which user can execute various kinds of traffic simulations. The simulator would be helpful to predict a traffic when we introduce a road pricing scheme, to predict a future traffic in an aging society with fewer children, to estimate the vehicles’ carbon-dioxide emissions in a city. Our paper is the first to simulate the traffic of millions of vehicles with completely agent-based. All the previous work simulate vehicles up to 100 thousands with incompletely agent-based. The rest of the paper is organized as follows. We overview MASMITS in the next section and explain how we model the driver behavior in the successive section. In the following section, we provide simulation results of MASMITS in two road network cases. The last section concludes our paper. Simulator Design In this section, we describe the design of the metropolitanarea scale traffic simulator, MASMITS. Firstly in the next subsection, we overview MASMITS. Then we give a description of the architecture in the successive subsection and how MASMITS works in the following subsection. Simulator Overview Our simulator is designed to be an experimental environment for simulating large city traffic with multiple driver behaviors. The strength of our traffic simulator is three fold. Firstly, the simulator can handle large number of agents, up to million drivers, which enables users to simulate the traffic of metropolitan in microscopic. This enables users to simulate the traffic flow in metropolitan area in microscopic with millions of vehicles on a single PC. The second point is that the interface between the simulation space and agent space is clearly defined. Users can model and develop their own driver model easily. The simulator space provides current states of traffic and the alignment of roads to driver agents in the agent space. The data includes current speed and positions for vehicles, the distance between cars, the curvature and gradient of road on which the specific vehicle is running. Thus all what driver model implementer to do is to model how each driver react by these provided data. The driver models for each driver are easily imported to the traffic simulator and this enables users to simulate traffic with their own driver model effectively. The third point is that the simulator is designed as microscopic model, which means that the movements of whole vehicle is decided by the logic of each vehicle agent. This is the main difference of our simulator to the former simulator is that the in our simulator, every driver behavior is modeled in microscopic. The former simulator, for example, models traffic congestion as macro phenomena and route choice of individual drivers as micro behavior (Yamashita et al. 2005). In this eclectic approach, the speed of each vehicle is determined by the road parameter independently of drivers intension. While these ”rigid” models are effective in verifying the routing approach, these models are inadequate in other applications since they lack of the diversified and complex nature of traffic. Simulator Architecture Figure 1 shows the architecture overview of the MASMITS. MASMITS consists of two spaces, the Agent space and the !