Future Plans

Activity generation

The above results use traditional origin-destination tables for demand generation. We intend to move our investigations to activity-based demand generation. Our plan for this is to start with a synthetic population, for example generated by Iterative Proportional Fitting (6). Next, by some method, activity patterns and primary activity locations are allocated to that synthetic population. This can, for example, be achieved by randomly drawing from census micro-sample data (6). With this, i.e. a synthetic population with activity patterns and primary activity locations given, the iterations are started. One strategy module will generate locations for secondary activities, another strategy module will generate activity timings, yet another strategy module will generate routes, and the agent database will maintain different plans, and evaluate them via the mobility simulation. First results of this, with activity timing and routes inside the feedback loop, were already successful (3,49), and a prototype for location choice also exists (37).

Message-based modules

It was explained earlier that a computational architecture for real-world multi-agent simulations should consist of at least two conceptual parts: the module for the simulation of the physical system, and the strategy generation module(s). As long as one remains within the framework of day-to-day learning, these modules are called sequentially, and it is possible to exchange information between them by a slow technology, for example via files. For ITS however, it is necessary to include within-day replanning into the simulation system. This implies that the simulation of the physical system needs to remain in permanent contact with the module(s) that compute(s) the strategies.

An implementation that achieves this and also maintains computational efficiency is to keep the strategy modules completely separate from the physical simulation, and to couple them via messages. More specifically, the strategy modules would send the agent strategies to the mobility simulation, which would attempt to execute them. The mobility simulation would send the events back to the strategy modules. In intuitive sketch of this is Fig. 7. Further details can be found in Nagel and Marchal (40) and Gloor et al. (28).

Figure 7: Virtual Reality Representation of Simulated Traffic in Portland/Oregon. Including visualization of message-based within-day replanning.