International Choice Modelling Conference, International Choice Modelling Conference 2015

Font Size: 
Revisiting the Route Choice Problem: A Multi-Level Modeling Framework for Route Choice Analysis
Evanthia Kazagli, Michel Bierlaire

Last modified: 11 May 2015

Abstract


Context and motivation The use of random utility models for route choice analysis involves challenges stemming from the high requirements in data and data processing, the physical overlap of paths, and the large size of the choice set. These factors significantly increase the complexity of the models. Given the complexity of the route choice problem, we aim at simplifying the models and their application in large networks.

The modeling framework proposed in this paper is based on a new approach in the way that the route choices are represented and modeled. The conventional representation and modeling approach is based on path alternatives constructed as link-by-link sequences on the network. This approach entails a very large number of possible paths connecting a given origin and destination (OD), and high correlation among the alternative paths. In this work, we claim that a path is solely the manifestation of the route choice, i.e. the way the traveler implements her decision to take a specific route, and we replace the paths with more aggregate elements that we denote as mental representation items. This key feature of the framework allows us to reduce the complexity of the model and at the same time is more behaviorally realistic.

Hypothesis Our behavioral hypothesis is that route choice takes place in a higher conceptual level that can be supported by an aggregate route representation (ARR). In order to accommodate the ARR in the model we introduce the concept of Mental Representation Item (MRI). A hierarchical ordering of the MRIs is assumed based on varying levels of detail in the representation. An MRI in the highest aggregation level can be the city center, a highway, or a bridge. The lowest level of aggregation corresponds to the conventional path representation.

Methodology Within the framework, route choice can be approached and analyzed in each of the layers of the underlying hierarchy. The definition of the layers is driven by modeling considerations controlling for the trade-offs between complexity, tractability and realism. It is important to note that all layers refer to the same choice. The difference lies in the level of aggregation in the representation. Each layer is characterized by a choice set of MRIs, denoted as Cl. Under the ARR assumption Cl is independent of the OD and common for all travelers. As an example, the most aggregate choice set could be Cl = {go through the city center, avoid the city center}. Subsequently, there is no need for choice set generation and it is solely the attributes of the alternative MRIs that are OD specific. Empirical evidence of such choice sets has been obtained by asking people from different cities (Athens, Stockholm, Lausanne) to state their route options from home to work.

The first step for the operationalization of this framework is the definition of the MRIs. The second step concerns the consistency of the results from the different layers. In this paper, we focus on the first step and we develop a simple procedure to generate the attributes of the MRIs in the most aggregate representation level in order to test the methodology.

Case study and results We use the network of Borlänge[1] in Sweden, for which map-matched trajectories of 24 private vehicles are available, as a case study. After examining the Borlänge network we identify a relevant aggregate choice set consisting of three alternatives Cl = {go through the city center, go around the city center, avoid the city center}.

Each observed path is assigned to one of the three alternatives following simple criteria. That is, if the path traverses links that are enclosed by the identified perimeter of the city center it is assigned to the first alternative; if the path traverses links only on, and not inside the perimeter, it is assigned to the second alternative; and finally if the path does not traverse any links related to the core or the perimeter of the city center it is assigned to the third alternative. This way each observed path is replaced by an MRI.

With respect to the definition of the alternatives we follow a deterministic approach that assumes a representative path. The representative path is the shortest path satisfying the criteria in order to belong to the corresponding MRI. For the generation of this path we use a simple procedure based on a gateway shortest path approach. The attributes of the corresponding MRIs are generated based on the representative paths.

Under this aggregation we reasonably assume that the alternatives are not correlated and we use a multinomial logit model with three alternatives to analyze route choice. The attributes that have been tested are the travel time, length, number of left turns, and number of intersections. In addition, it is possible to estimate alternative specific constants –something that is not feasible with the conventional approach.

The estimation results are consistent, the parameters have the expected signs and they are significant. We are currently working on improving this specification by including OD specific attractors for each MRI –e.g. a shopping mall that someone would encounter on his way.

The next step of the work consists in the assessment of the performance of the presented approach for traffic assignment.

Conclusion The example presented in this paper illustrates the potential of the proposed approach to simplify the route choice problem by reducing i) the size of the choice set and avoiding path generation and sampling of alternatives, and ii) the correlation among the alternatives. An important step for the validation of the proposed framework consists in its use for traffic assignment, and comparison with the conventional route choice models, in order to identify the trade-offs between the full and the simplified choice set with respect to parameter estimates and predictive capabilities.

References

Axhausen, K., Schönfelder, S., Wolf, J., Oliveira, M. and Samaga, U. (2003). 80 weeks of GPS traces: Approaches to enriching the trip information., Transportation Research Record: Journal of the Transportation Research Board 1870: 46–54.

Frejinger, E. and Bierlaire, M. (2007). Capturing correlation with subnetworks in route choice models, Transportation Research Part B: Methodological 41(3): 363–378.


[1] We refer to Frejinger and Bierlaire (2007) and Axhausen et al. (2003) for a description of the Borlänge GPS dataset.


Conference registration is required in order to view papers.