Summary / Conclusions

This paper presents a framework for large scale multi-agent simulations of travel behavior. The immediate goal is a replacement of the traditional four-step process for transportation planning; the longer term goal is to have an agent-based system for all aspects of urban and regional planning. The core ideas of the system are:

• One should separate the simulation of the physical world from the simulation of the mental world of the agents. This is reflected by a design that strictly separates those two worlds. A mental/strategic layer generates strategies (``plans''); these plans are executed by a physical layer; and then strategies are adapted based on how they performed in the physical layer. In the specific case of transportation simulations, the physical layer is also called the traffic (micro-)simulation, or the mobility simulation.

• Agents should memorize more than one strategy. This gives them the option to try out all strategies multiple times, and thus to obtain average scores for them (``exploration''). When agents do not explore, then they exploit, by selecting strategies based on the score, possibly using some behaviorally justified model such as multinomial logit. From time to time, agents should also add new strategies to their repertoire, or delete strategies with low scores. All this functionality is implemented by a so-called agent database.

• As said above, agents from time to time add new strategies to their repertoire. These strategies need to be generated by some method. In our approach, the methods that generate new strategies can be completely independent modules. They read an XML file that specifies which pieces of an agent's plan are already known, and the external module adds or changes pieces. For example, one module could generate activity patterns, another one activity locations, a third one activity times, and a fourth one routes. The framework will call all those modules one by one and in the end have a completely new plan. The XML plans file uses exactly the same format for all those exchanges; and the same format is again used to send plans to the physical simulation. This means that we have only one file format specifying agent plans.

• Performance information from the physical simulation is communicated back via an events file. Events are output every time something relevant happens, such as an agent leaving an activity, or entering a link. In general, the physical simulation performs no data aggregation at all; in consequence, something like link travel time needs to be constructed from link entry and link exit events. The two big advantages of this approach are: (1) It is straightforward to retrofit those outputs to any existing agent-based mobility simulation, since there is no data aggregation either along space or along time. Any event is just a simple print command inside the mobility simulation. (2) External strategy generation modules can also read this information and use it as they like. For example, an external module building a mental map for one specific agent will use the link entry/exit information in completely different ways than the routing module. If the simulation would aggregate link entry/exit information into link travel times, adding such a mental map module would no longer be possible.

• Strategies are full 24-hour day plans, and are also evaluated as such. The strategy evaluation is based on the events, and therefore on actual individual performance of each agent. The evaluation is therefore immune against errors introduced by aggregation. One obvious candidate for the scoring function is the conventional utility function, but other systems such as aspect theory can be used. Since all agents are individually represented, it is no problem to couple those evaluations to individual attributes, such as gender or income.

• The approach is completely transparent toward parallel computing. Having a parallel mobility simulation needs some programming effort, but is conceptually straightforward since methods from the simulations of other physical systems can be applied. And making the agent database or the external strategy generation modules parallel is completely trivial as long as agents do not interact: One simply distributes agents (or possibly households) across CPUs. Only when interactions beyond the household level become important will more advanced computing techniques become necessary.

• The implementation was tested with three scenarios, one testing route equilibration, one testing times choice, and and one testing route equilibration in conjunction with time choice. Route equilibration works as expected, in the sense that the system ends up with using nine equivalent routes equivalently. Also time choice works as expected, in the sense that agents adjust their daily timings in a plausible way.

• A curious aspect about the time choice test is that the external module that generates alternative timing schemes has no ``intelligence'' at all: Instead, it just randomly mutates existing schedules of the agents. The interpretation of the new schedules is entirely done by the agent database and its scoring function. This means in practice that the correctness specifications toward external modules are significantly relaxed: In essence, it is sufficient if external modules generate meaningful solutions plus some variability.