Introduction

For urban planning, it would be useful to have a computational tool that evaluates transportation consequences of urban evolution scenarios. The anticipated urban structure, including anticipated demographic data and anticipated transportation infrastructure, could then be fed into a computer, and the computer would calculate the resulting traffic. Maybe the tool would have a virtual reality component, enabling the user to ``fly'' to, say, a critical intersection and count the number of cars during the rush hour (see Fig. 1). Such a tool should include the effects of ``induced'' traffic, i.e. the fact that lower congestion levels encourage people to do more travel; the tool should be able to say something about the variability of traffic; and it should be capable of incorporating new technology, such as tele-commuting or telematics systems.

The problem of induced traffic makes clear that transportation is not a simple infrastructure problem, where to find a good or optimal solution to move a given demand, but a problem where individual people's values and preferences play an enormously important role. In consequence, any solution technique needs to be able to represent aspects of human decision-making. The problem thus becomes as much a problem of (computational) social science as one of engineering.

The TRANSIMS (TRansportation ANalysis and SIMulation System) project at Los Alamos National Laboratory [] is an attempt to build such a tool. The key to the TRANSIMS design is that it is completely microscopic, which means that it keeps track of individual travelers throughout its modules. Similarly, elements of the transportation infrastructure, such as intersections, traffic lights, turn pockets, etc., are represented microscopically.

In TRANSIMS, each traveler is a computational agent. Agents make plans about what to do during a day - in order to get from one activity location to another, agents can, for example, walk, use bicycles, drive cars, or use busses. Eventually, all plans are simultaneously executed in a micro-simulation of the transportation system.

In principle, this leads to a straightforward simulation approach (see Fig. 2): Derive synthetic households from demographic data and locate them on the network; use the demographic information together with land use information to derive activities (working, sleeping, eating, shopping, etc.) and activity locations for each household member; and let agents decide about mode and routing for their transportation. So far, all these are plans, i.e. intentions of the simulated individuals. These plans can then all be fed into a realistic transportation micro-simulation, which can be used as the basis for analysis, such as emissions calculations.

The advantage of such a microscopic approach is that, at least conceptually, it can be made arbitrarily realistic. This makes it possible to include dynamic effects such as queue spillover, which means that congested traffic can spill back across intersections, and which are sometimes hard to represent in traditional methods. It also makes it possible to include new and perhaps unanticipated technology at a later time. For example, the whole architecture of ITS (Intelligent Transportation Systems) can be mirrored by a careful software implementation.

Yet, there are also several disadvantages, some of them being:

(i) Size of the problem: Metropolitan regions typically consist of several millions of travelers. Executing a second-by-second transportation micro-simulation on a problem this size within reasonable computing time is only possible with the use of advanced statistical and computational techniques.

(ii) Behavioral foundation: We are far from understanding human behavior. For that reason alone, we are unable to predict the behavior of individual travelers. However, there is a chance that the macroscopic (emergent) behavior that is generated by thousands or ten-thousands of interacting individuals is considerably more robust than the behavior of an individual agent. This would be similar to Statistical Physics, where the trajectory of a single particle is unpredictable, yet, useful macroscopic properties of gases such as equations of state can still be derived.

(iii) Consistency problem: The approach outlined above is not as straightforward as it sounds because the plans depend on expectations about traffic conditions. For example, if a person expects congestion, he or she may make different plans than when no congestion is expected. Yet, congestion occurs only when plans interact during their simultaneous execution. In short, plans depend on congestion, but congestion depends on plans. - This logical deadlock is not unknown in economic theory and is traditionally overcome by the assumption of rational agents. Both with and without the assumption of rationality, this problem of consistency between plans and micro-simulation makes the computational challenge even bigger.

(iv) Robustness: Any approach to a problem needs to have reproducibility of the results under a wide enough range of changes, or otherwise the results are useless for practical purposes.

This paper concentrates on the computational problems, and what we have done to overcome them. We start by defining the general problem of dynamic traffic assignment (Sec. 2). Sec. 3 describes the application framework, which essentially says that we need to iterate between route planning and micro-simulation. Secs. 4 and 5 explain the route planner module and the micro-simulation module, with special emphasis on our parallel implementation. Secs. 6 and 7 report actual performance results from simulations in Dallas/Texas and in Portland/Oregon. The paper is concluded by a short summary.

  

Figure 1: Virtual reality transportation system representation.

  

Figure 2: TRANSIMS design