This paper describes aspects of large scale multi-agent simulations (MASim) for traffic simulations in general, and the evaluation of ITS systems in particular. A first important starting point is that such simulation systems consist of two components: the simulation of the physical system (mobility simulation), and the simulation of the strategy generation. Both modules are rather different, and thus need rather different techniques in terms of modeling and implementation.
It is important to note that ITS devices can be treated as agents in a similar way as the travelers. A strategy for an ITS device could for example be a message on a variable message sign, or a signal timing plan, and both could be computed by a traffic management center and then sent to that device for execution in the mobility simulation.
Besides the two components ``mobility simulation'' and ``strategy generation'', there also needs to be an implementation of agent learning. Single-agent learning can be understood from a behavioral perspective, treated elsewhere in this workshop, or from a computer science perspective, where it touches upon Artificial Intelligence and Machine Learning. Importantly, when several agents learn together, then the whole learning system can be described as a dynamical system. As is well known, dynamical systems go toward attractors, which can be fixed points, periodic, or chaotic. There is nothing in our knowledge that tells us where a simulation of a learning traffic system will go, or where the real system will go. In addition, issues of ``learning speed'' matter, in particular when there are several entities (such as users and Traffic Management Center) which learn simultaneously.
MASim looks like the perfect technology to evaluate ITS systems, since is is possible to implement individual behavioral rules for each individual traveler, thus allowing for a differentiated and segmented response of the traveler population. Also, traffic management operations can be included into the simulation framework in a straightforward way: each variable entity, such as a traffic signal or a variable message sign, just becomes a ``technical'' agent by itself, that follows a plan given to it by the traffic management center.
In order to explore large systems, significant computational speed is necessary. Issues of computational speed are discussed, demonstrating that it is possible to study several millions of travelers with computation times on the order of a day.
Unfortunately, the technology that enables such agent-based large scale simulations does not allow within-day replanning, i.e. the capability of the agents to change their plan while en-route. This is, however, clearly necessary for the evaluation of ITS. Therefore, future plans include coupling the different simulation modules via messages. This would not only further increase computational speed, but it would also allow the direct implementation of within-day learning.