Human transportation has physical, engineering, and socio-economic aspects. This last aspect means that any simulation of human transportation systems will include elements of adaptation, learning, and individual planning. In terms of computerization, these aspects can be much better described by rules which are applied to individual entities than by equations which are applied to aggregated fields. In consequence a rule-based multi-agent simulation is a promising method for transportation simulations (and for socio-economic simulations in general). By a ``multi-agent'' simulation we mean a microscopic simulation that models the behavior of each traveler, or agent, within the transportation system as an individual, rather than aggregating their behavior in some way. These agents are intelligent, which means that they have strategic, long-term goals. They also have internal representations of the world around them which they use to reach these goals. Adding the term ``rule-based'' indicates that the behavior of the agents is determined by sets of rules instead of equations. Thus, a rule-based multi-agent simulation of a transportation system will apply to each agent individually. Such a simulation differs significantly from a microscopic simulation of, say, molecular dynamics, because unlike molecules, two ``traveler'' particles (agents) in identical situations within a transportation simulation will in general make different decisions.
Such rule-based multi-agent simulations run well on current workstations and they can be distributed on parallel computers of the type ``networks of coupled workstations.'' Since these simulations do not run efficiently on traditional supercomputers (e.g. Cray), the jump in computational capability over the last decade has had a greater impact on the performance of multi-agent simulations than for, say, computational fluid-dynamics, which also worked well on traditional supercomputers. In practical terms, this means that we are now able to run microscopic simulations of large metropolitan regions with more than 10 million travelers. These simulations are even fast enough to run them many times in sequence, which is necessary to emulate the day-to-day dynamics of human learning, for example in reaction to congestion.
In order to demonstrate this capability and also in order to gain practical experience with such a simulation system, we are currently implementing a 24-hour microscopic transportation simulation of all of Switzerland. Switzerland has 7.2 million inhabitants. Assuming 3 to 3.5 trips per person per day, this will result in about 20-25 million trips. This number includes pedestrian trips (like walking to lunch), trips by public transit, freight traffic, etc. The number of car trips on a typical weekday in Switzerland is currently about 5 million (see Vrtic (2001) for where the data comes from). The goal of our study is twofold:
This paper gives a report on the current status. Section 2 describes the simulation modules and how they were used for the purposes of this study. Section 3 describes the input data, i.e. the underlying network and the demand generation. Besides ``normal'' demand, we also describe one where 50000 travelers travel from random starting points within Switzerland to the Ticino, which is the southern part of Switzerland. We use this second scenario as a plausibility test for routing and feedback. This is followed by Sect. 4, which describes some results and Sect. 5, which describes issues related to computational performance of the parallel micro-simulation. The paper ends with a discussion and a summary.