A possible goal for transportation and regional planning is to design the transportation or regional systems so that the people who use them will be happy. Yet, happiness is difficult to define in a way which is useful for engineering. For most people, a certain level of economic performance is probably a pre-requisite for happiness, and thus economic indicators are important factors in the design. Other important indicators might include noise, pollution, safety, access to a variety of destinations, or even something as intangible as beauty. These indicators may have different impacts on the happiness of different users of the above-mentioned systems, so design of these systems should somehow account for individuality of the users.
The traditional quantitative method for transportation planning is the four-step process (e.g. Sheffi, 1985). The four steps are trip generation, trip distribution, mode choice, and route assignment. In the first two steps, the origins (e.g. residential areas) and destinations (e.g. workplaces) are determined and then connected. The typical result of these steps is an origin-destination matrix, which gives, for each zone in the simulated region, the number of trips to each other zone in a given time slice. The last two steps then determine which mode and which path these trips take.
The main disadvantage of this technology is that all relation to individual people is severed in the process. This has consequences both on the level of modeling and on the level of analysis. For example, on the level of modeling one would expect people's income or distance to a bus stop to be important for mode choice, but the four step method does not take these into account. Similarly, the fact that individual travelers' characteristics are not available makes any analysis based on individual people's characteristics impossible.
An alternative, which will be discussed in this paper, is to maintain the individuality of travelers throughout the modeling process. Doing this is now possible with ever more powerful computers, and the new computational tools that have emerged with them. The computational tools that are particularly relevant with respect to this paper are the methods of Complex Systems, which includes for example Cellular Automata, Genetic and other Evolutionary Algorithms, or Multi-Agent Simulation.
The general idea of such an approach is easy to explain: In a multi-agent transportation simulation, travelers are represented as individual ``agents'' who make independent decisions about their actions. As will be explained later, these decisions range from short-term decisions about acceleration or lane changing all the way to decisions about the planning of daily activities, and later versions of such a modeling system are bound to also include long-term decisions about, say, residential choice.
Note that such an approach still allows planners to extract more conventional aggregated quantities, such as, for example, the fraction of travelers who chose public transit. In contrast to more traditional modeling approaches, in the agent-based system this number will be generated by looking at each traveler individually, and carefully weighting all her/his different options, taking into account for example her/his income and her/his distance to the next bus stop. This is important since it is expected that such details play an important role with respect to choice behavior. In consequence, the aggregated number will transmit the same information as before, but it is now based on a much more microscopic evaluation than before. The same argument applies to all kinds of transportation system analysis. As further examples, consider an analysis of who is affected by noise or air pollution (children!), who is contributing to a traffic jam (do they have an alternative?), or who should pay for public transit (which frees up roads for car users).
This paper introduces the most important aspects of such a multi-agent simulation system for transportation planning, and in fact for all kinds of mobility simulations. Such a simulation system consists of many modules which need to work together. On the highest level, one can distinguish the following classes of modules: First, the movement simulation itself, which moves the agents through the system, and which is concerned with the physical aspects of the system such as volume exclusion (two travelers/cars cannot be at the same location at the same time) or limits on acceleration/braking. This is treated in Sec. 2. Second, tactical/strategic modules model higher-level decision-making such as route choice, generation of daily plans, or residential choice. This is described in Sec. 3. Finally, there needs to be a learning framework, which models how how agents can learn to improve their decisions as the simulation progresses (Sec. 4). After the discussion of these three conceptual pieces of the simulation system, Sec. 5 describes a scenario of real-world rush hour traffic in Switzerland that was executed by our multi-agent transportation simulation system, and results from that simulation. Section 6 presents some issues with computational performance of multi-agent simulation modules. The paper is concluded by a summary.