A Review of Theoretical and Practical Issues in Microsimulating Transport Demand
Eveline Helder, Significance, Michiel De Bok, Significance, Gerard De Jong, Significance
We review the simulation error in current microsimulation models for transport demand. We formulate guidelines for application in the agent-based strategic passenger transport models for Flanders.
With the availability of increased computing power and large databases, microsimulation is becoming more popular as a method in practical transport planning models. Microsimulation often involves the application of Monte Carlo simulation which introduces simulation error into the model outcomes. Conventional trip-based demand models do not even distinguish behaviour of individual agents and operate on the basis of the number of trips between two zones. The more advanced disaggregate tour-based models in application usually determine choice probabilities for specific market segments (person-types) and sum these (“sample enumeration”) to zone-to-zone matrices. In contrast, a microsimulation approach for transport demand predicts the behaviour of a simulated agent choosing a tour frequency, destination, mode and/or time period alternative, by selecting a single discrete alternative from the cumulative distribution function. The selection is based on the outcome of a random draw (usually from the uniform distribution between 0 and 1). Hence, using a different seed for this random process will lead to different output matrices. This randomness in the results leads to some practical issues for the implementation of microsimulation models in transport demand models.
This research is undertaken as part of the development of the new agent-based strategic passenger transport models for Flanders. The demand model is based on a series of tour-based travel demand models that were estimated on the Flemish travel surveys. The actual microsimulation implementation of the estimated models is described in a companion paper (Verlinden, Puttemans, de Bok, de Jong and Helder).
The objective of this more general and theoretical paper is twofold:
1. identify and discuss theoretical and practical issues that follow from microsimulating transport demand;
2. formulate practical guidelines for the implementation of microsimulation in transport demand models.
For this purpose, feedback is used from a selection of experts in the field of operational microsimulation models whom we asked for their experience on this topic. On the basis of this feedback and our own experiences so far, we will address the following issues that we consider relevant for the implementation of microsimulation models for transport demand models: simulation error, the required number of replications and the implications of ‘randomness’ in individual results.
Simulation error (or simulation variance) is a result of the Monte Carlo simulation. Various measures exist that can be used to quantify simulation error, such as the standard deviation on the outcomes. The simulation error depends on the number of replications that are generated for the outcome distribution and on the predicted value: the smaller the output value, the larger the simulation error, relative to the outcome. It is shown how the simulation error (measured as the standard deviation of the outcome) can be approximated a-priori. This result can be used to calculate a 95% confidence interval of the outcome (reflecting simulation error). Transport demand models are composed of a series of submodels, each producing simulation error that is propagated in the model application. A methodology is formulated that can be used to quantify the simulation error on a series of submodels.
In practical applications, the required number of replications will depend on the size of the impact of a policy measure and the required level of detail in the evaluation criteria. The number of replications in a model application determines the reliability of a prediction. This reliability depends on the dimensions of the evaluation indicator (e.g. by zone, purpose, person-type). In general: smaller predictions have a lower reliability: so the smaller the project effect, the more replications will be needed. We will propose rules-of-thumb or standard procedures to set the required number of replications in practical model applications, as well as rules to obtain the final “best” prediction from a series of draws (simulated working days) with different seeds
The evaluation of output indicators at the microlevel is appealing for story-telling, but also holds a risk of damaging the face-validity of the model, when randomness may causes counter-intuitive results. Recommendations are formulated on how to, or how not to, use results at microlevel.
Finally, the paper formulates a general strategy for dealing with microsimulation in a strategic transport demand model such as the one for for Flanders. This strategy consists of:
- a procedure to determine the number of replications required for typical applications;
- statistical routines to calculate confidence intervals for intermediate and final model results;
- recommendations on the presentation of final results.
Association for European Transport