The main research question of this thesis was: How can the assignment of cyclists in a four- stage model realistically be modelled for use in the Netherlands?. This question was
answered by answering the following subquestions:
1. What are the most relevant factors in the route choice in the Netherlands? 2. What methods do exist for the assignment of cyclists in a four-stage model? 3. What is the most useful method for assignment in the Netherlands?
4. What combination of variables does describe the route choice in the Netherlands best? 5. How well does the all-or-nothing assignment perform in the Netherlands?
6. To which extent is the proposed algorithm an improvement?
Many factors seem to influence the route choice of cyclists. Previous research shows a broad range of factors, from aspects in the environment to road characteristics and
personal factors. Finding the most relevant factors was not clear-cut. Nearly all literature is focussed on cities in other areas and/or is less applicable for an assignment because it is a stated-preference study. Experts at Goudappel Coffeng were asked about their opinion on the most important factors in the Netherlands and the locations where previous research took place, were compared. Nearly all sources mentioned distance as one of the most important factors together with travel time. The road type was important in most research, but in one study it was not significant. Intensities of motorised traffic were significant in most relevant studies, but no data is available about that in a bike assignment, except indirectly via the road type. The literature was less clear about for example the influence of the land use around the roads and therefore it was not selected. The selected most relevant factors were distance, time, type of intersection and road types as there seemed to be more agreement on the relevance and because of the availability of data.
The second subquestion focusses on the algorithm which should assign the cyclists using the factors found in the previous subquestion. Cyclists are currently assigned using the all- or-nothing algorithm. In literature, other options were explored, like logit and probit choice models and a user-equilibrium assignment with congestion included. Logit is most used in literature and is able to find multiple routes in contrast with all-or-nothing. A downside is the handling of overlapping routes as independent which they are not. Probit has similar
abilities, but is more statistically accurate than logit. It needs Monte Carlo simulation for the calculation of the probabilities, but that takes much time and is therefore a disadvantage. The user-equilibrium assignment is based on the speed, which is on its turn based on road characteristics, car intensities and cycle intensities. To overcome some of the
disadvantages of logit and probit a path-size factor can be added to the logit model to correct for the overlaps without the need for simulation.
The most relevant algorithm was selected on the basis of criteria like running speed, ability to assign different routes and the possibility to include the factors selected in the second subquestion. The logit model with path-size factor scored best on these criteria and was therefore selected as algorithm. The probit model was not chosen because of the
about the method and the difficulty in applying the factors. Logit needs routes to choose from and to do so, a Dijkstra’s shortest route algorithm was used with Monte Carlo simulation to vary the costs and create multiple routes.
The fourth subquestion was answered by implementing the logit and route generation algorithm in OmniTRANS and calibrating the parameters to find the best fitting combination of values for those. First of all, the route generation algorithm was calibrated on a less detailed ’s-Hertogenbosch network for testing purposes and subsequently on Tilburg. Generating 12 to 20 routes and using 0,15 as variance provided the best balance between running time, storage and quality of the generated routes. The assignment was only calibrated on the Tilburg network. Two combinations scored best: one with the least large deviations from counts and one with the most small deviations from counts. These settings both give lower costs to routes with separate cycle paths and higher costs to routes with traffic lights. Using the best setting 44,7% of the counts scored a t-value lower than 3,5, which is a small improvement over the all-or-nothing assignment.
The all-or-nothing assignment did not score well in these networks. In Tilburg, only 42,4% of the counts had a t-value lower than 3,5 and in The Hague only 43,2%. In both networks around 36% of the counts scored a t-value higher than 4,5. Furthermore, the all-or-nothing assignment does not represent cycle route choice well: it is only able to assign cyclists to one route between an origin and a destination, while cyclists choose different routes, as the literature and experts mentioned. Furthermore, in most cases more than one factor
determines the route, not only time or distance.
The algorithm was compared with an all-or-nothing assignment applied to the Tilburg network to assess the improvement. The improvement seemed small, the number of good counts increased with 2,4 percentage points using the t-test and it did not improve using the GEH-statistic. The improvement was also assessed on the network of The Hague, Delft and surrounding areas. On this network the logit assignment had an increment of 3,6 percentage points of counts in the best category. The number of counts in the worst category decreased with 0,7 percentage points. The GEH-statistic improvement was bigger: an increase of 6,2 percentage points in the proportion of counts in the best
category. The logit assignment can therefore be seen as a slight improvement to the all-or- nothing assignment. More positive is that cyclists were assigned to routes with more cycle facilities and less intersections, but the number are not correct in all instances.
All in all, the logit assignment seems to be a method that is an improvement of the all-or- nothing assignment for cyclists. The used parameters resulted in a small improvement, but there is potential for better scoring assignments using different and more factors. The algorithm performed even better during validation and seems therefore to work in different areas within the Netherlands, which has not been tried before. To increase the accuracy of the assignment, more research can help collecting more data about cyclists’ routes in the Netherlands and optimising the algorithm by changing parameter values and adding additional factors.