Large scale multi-agent transportation simulations

In a multi-agent transportation simulation, each traveler is represented individually. Such simulation consist of at least the following modules: - Activity generation. For each traveler in the simulation, a complete 24-hour day-plan is generated, with each major activity (sleep, eat, work, shop, drink beer), their times, and their locations. - Modal and route choice. For each traveler in the simulation, the mode of transportation and the actual routes are computed. - The Traffic simulation itself. In this module, the travelers are moved through the system, via the transportation mode they have chosen. A reasonably realistic traffic dynamics is necessary to include dynamic effects such as queue built-up. - Learning and feedback. In order to find solutions which are consistent between the modules ("congestion is a result of plans, but plans are made in anticipation of congestion"), a standard relaxation technique is used. This technique has similarities to day-to-day human learning and can also be interpreted that way. It is clear that further modules need to be added, such as for housing and land use, but also for freight traffic. The important point of doing rule-based microscopic simulations is that it is possible to experiment with arbitrary behavioral rules, going all the way from simple learning heuristics to rational agent That is, one is not bound by, e.g., mathematical constraints. It is for example possible to construct, for each individual agent, a large set of plans ("strategies") and have the agent select between these strategies, based on past performance, or construct a new strategy. This allows, for example, to evaluate performance according to individual preferences. It also allows to have, for each agent, an only partial knowledge of the world, which may be different for each agent, and may be changed via exploration ("mental maps"). Using advanced computational methods, in particular parallel computing, it is now possible to do this for large metropolitan areas with 10 million inhabitants or more. We are currently working on such a simulation of all of Switzerland. Our focus is on a computationally efficient implementation of the agent-based representation, which means that we in fact represent each agent with an individual set of plans as explained above. We use a data base to store the agent's strategies, then load them into the simulation modules as required, and feed back individual performance measures into the data base. This approach allows that additional modules can be coupled easily, and without destroying computational performance. Since the above only models day-to-day replanning, we also look at within-day replanning, which means that travelers can change plans during travel. In particular, we look at efficient distributed implementations of this. It turns out that computational efficiency is closely related to the real-world mechanics of the distributed intelligence inherent in the real world system.