State of the art

No simulation package currently integrates all the aspects that are discussed. TRANSIMS [40] comes from the transportation planning side and is maybe the most advanced in terms of using many of the concepts. The TRANSIMS research program is reaching completion in 2002, with a full-scale simulation of a scenario in Portland/Oregon, with a network of 200000 links and several million travelers. TRANSIMS uses an agent-based approach to how the travelers make decision, including an agent data base and a selector which selects agents for replanning. TRANSIMS does however not implement any of the ``within-day replanning'' aspects discussed in Sec. 4.3. We ourselves are in the process of using TRANSIMS for a full-scale simulation of all of Switzerland [41]. We also have prototypes for the ``truly distributed intelligence'' as discussed above [24], and for genetic algorithm usage for plans generation [42]. DYNAMIT [43] and DYNASMART [44], originally started as transportation simulation tools for the evaluation of ITS (Intelligent Transportation System) Technology, also advance into the area of transportation planning by the addition of the demand generation modules. METROPOLIS [45] is a package designed to replace static assignment by a simulation-based but very simple dynamic approach. It allows the user to specify arbitrary link-cost functions but in its current version is does not allow the queue build-up which is important for congested systems. The strength of METROPOLIS lies in the self-consistent computation of departure time choice. Very few projects use individual plans. Instead, they use shortest-path trees as described in Sec. 4.2. A collection of articles about regional transportation simulation models can be found in [46].

Thus, for real world implementations, there is still a long way to go until the agent-based approach is truly implemented, let alone tested. As an example of the few existing comparisons to real world data, Fig. 10 shows such a comparison for a Portland (Oregon) scenario done within the TRANSIMS project. The compared data are hourly flows, i.e. the number of cars crossing certain measurement locations during an hour. For both plots, the x-coordinate of a point is given by the field data value, while the y-coordinate is given by the model result. In consequence, the deviation from the diagonal is a measure of how much field data and simulation result disagree. Each point denotes a different measurement location. The left plot shows results of our simulation, while the right plot shows results from a model run done by the Portland transportation planning authority using more traditional technology. The result says that agent-based simulations in transportation currently are, in terms of the quality of the result, comparable to the more traditional technology; this statement can be quantified [47]. When interpreting this result, one should consider that our result was preliminary, with a much simplified micro-simulation and a much simplified demand generation, while on the other hand the Portland transportation authority has a reputation for excellent modeling work. For further details, see [47].

Figure 10: Comparison between field data and model results. LEFT: Our method. RIGHT: Portland transportation planning authority