Within the context of travel behavior research, many approaches look at individual travelers as the principal unit of research. This holds true for example for random utility models, which predict behavioral probabilities as a function of individual demographic and individual option-dependent characteristics, but also for psychological/rule-based approaches. Using these approaches for transportation policy will only be possible if those results can be put into a framework where it applies to real-world transportation-related questions. One viable method to achieve this are multi-agent simulations, which can treat several millions of individual agents within a consistent simulation framework. The agent-based approach has the advantage that it can directly incorporate behavioral research results; the same is not true for the 4-step process or any similar aggregated approach. It is thus possible to program a multi-agent simulation with the rules and results of travel behavior research and thus obtain a simulation system for policy forecasting that is based on the research results.
The first part of this paper explains how such a simulation system could be designed. There is increasing consensus in the Artificial Intelligence community that in order to understand intelligence it is necessary to have the mental simulations interact with the real world. This ``embodiment hypothesis'' is, in the context of travel behavior research, best put into practice by having the strategy generation separated from the mobility simulation, which receives the strategies of the agents, executes them, and feeds back information about the agent's actual experiences (such as incurred travel times). Besides strategy modules and a mobility simulation, a functional system needs initial/boundary conditions, and a learning method.
The second, considerably longer, part of this paper then discusses computational and implementation methods. Among the discussed areas are the application of computer science search methods including evolutionary algorithms to travel behavior modeling, or capabilities that databases could and should have to be even more useful. Special emphasis is given to the trade-offs between ease-of-programming, interoperability, and computational performance. In particular, we expect that in the future much progress will be made by coupling modules coming from different research groups, which will mean that codes using different programming languages running on different operating systems need to be coupled. The paper attempts to give some guidance plus pointers to relevant methods for these cases, and how they relate to performance-oriented parallel computing. The overall conclusion is that many of the necessary aspects of inter-operable software construction are just emerging. Thus, there is hope that the necessary methods will eventually become available, but at this point, many aspects are still rather experimental. In contrast, as long as modules run sequentially (that is, within-period replanning is excluded), very helpful standardized methods such as XML, CORBA, Java RMI, or MPI are now available.
The paper deliberately concentrates on fairly technical questions, since such aspects are potentially important for large scale implementations, and they are rarely discussed in sufficient detail. In particular, the computational methods part of the paper is meant as a compendium for someone considering to implement parts of such a simulation package. Overall, it is hoped that this paper stimulates discussion about how travel behavior research models can be implemented, and which features may be desirable to make them inter-operable.