Existing navigation systems help drivers to navigate through traffic and hence improve driving comfort. Past researches have concluded that the satisfaction level of the systems appears to be relatively low which can be related to the level of network familiarity. A reasonable approach to improve the user satisfaction is to personalize navigation systems by developing a route planning process which would be able to learn user preferences through interaction with the driver and enabling it to plan better routes for that driver in future operations. Within this interaction it is important to acquire knowledge on the preferences whilst minimizing efforts on the users, for example by discovering regularities when observing repeated route choice behavior.
The paper of Park et al. (2007) discusses various methods which could be applied to establish a user model for routing algorithms. It compares the discrete choice analysis models, based on utility maximization, with the non-parametric methods like data mining. Main characteristic of the latter is that the non-parametric methods do not require estimating any parameter by describing the distribution of variables and furthermore do not assume any particular form of functions.
Within the family of non-parametric models, two distinct algorithms can be distinguished which can be used to accommodate route choice rules discovered from the observed (revealed) travel behavior data. Firstly the Artificial Neural Network (ANN) which provides models of data relationships through highly interconnected, simulated “neurons”. These “neurons” accept inputs, apply weighting coefficients and feed their output to other “neurons” which continue the process through the network to the eventual output. The alternative, the Decision Tree Learning (DTL) algorithm uses a decision tree as a predictive model which maps observations about an item to predict the future target value of an item. The authors choose to apply the DTL algorithm although the ANN algorithm outperforms other machine learning methods. Reason is that the ANN methods utilizes a ‘black box’ and supplies output that is very difficult to interpret. Opposed to the ANN method, the DTL algorithm provides a more comprehensible model structure which simplifies the interpretation of the modeling results.
Main aim of the paper of Park et al. (2007) was to explore the advantages of the DTL algorithm for developing adaptive route guidance, within the study a decision tree is constructed by making use of the C4.5 algorithm. This specific algorithm has been widely applied, the classifying is straightforward and it is fairly simple to apply.
System architecture
Two route guidance architectures can be identified, firstly the ‘autonomous navigation’ in which the car functions largely independently. The car receives traffic information which should subsequently be evaluated by the on-board navigation device. The second architecture, the so called ‘supported navigation’, consists of navigation devices which receive pre-determined candidate paths which are generated by the traffic information centers and afterwards downloaded to the subscribed cars. To simplify the research the authors of Park et al. (2007) only focused on the route advices based on the autonomous system architecture.
100 | P a g e
The system design for adaptive personalized routing, as proposed in the research or Park et al. (2007), initiates with the generation of a set of feasible routes based on the available network data and traffic information. Subsequently the
predictive model assesses all candidate paths and selects the route which fits the user. This route can be subsequently presented to the driver and after reaching its destination the system determines the implicit reaction of the user by comparing the observed route with the suggested route. This approach, in which the passive user feedback (expressed as acceptance, rejection of deviation on-route) is used, enables the system to acquire knowledge of user preferences effectively without asking the drivers explicitly. If the user feedback is negative, the learning process is executed, leading to the updating of route selection routes.
The learning algorithm can be defined as ‘deducting knowledge from experiences with respect to some
classes of tasks and performance measure’; with respect to adaptive route guidance the following three elements can be described as:
A task is the search for a route which best fits driver’s preference using a decision tree;
A performance measure is the extent to which the route corresponds to the driver’s preference;
An experience comprises of personal, route and trip characteristics along with the route choice.
Experimental design
Within the paper of Park et al. (2007) experiments with the learning model were carried out to analyze its application and learning ability in the context of route choice. During the research it was difficult to observe actual route choice data, at the time the research was executed no methodology and resources were available to develop a framework to follow a large user group. As an alternative the route choice experiments were generated using the simulation program ‘ICNavS’. The first major step was to collect suitable data on route attributes and driver choices. Subsequently the best routes are determined by lexicographical rules and utility maximization rules described in Peeta and Yu (2005). After this pre- processing, a set of routes between the origin and destination is input to the learning model and a route is chosen by initial decision tree. If the result is different from the observed (revealed) choice, the decision tree is updated. In order to gain more accurate results of the adaptive learning algorithm and to facilitate the comparison of the results, three different sets of route selection routes were applied; two sets of lexicographic rules and one set of utility maximization.
Results
Various indicators were used to assess the performance of the various route choice algorithms. The first indicator was the number of updates of the data set, since the decision trees are updated whenever the actual choice of the driver is different from the predicted choice this measure is valid. A second efficient method to represent the predictive accuracy of
Learning process Pre-process
Initial data with route attributes and choice results Constructing an inital route choice model Choosing a destination Generating a maximally disjoint
path set Deciding the most likely route
with the model Suggesting the best route Collecting user feedback Adding the new case in the
additional data set
Updating the route choice model User feedback? Exit Positive Negative
101 | P a g e
1
0
1
the various algorithms was to plot the percentage of predictions that correspond to actualchoices made by the user over the period of time that the system has used. For the purpose of analyzing how well the models accommodate the route selection rules, error rates of each tree were computed by dividing the number of incorrect choice decisions predicted by the current tree by the size of the data set for constructing the current model so far. This result shows how well the adaptive routing algorithm is learning from past errors.
From all the combined results it became apparent that the DTL approach seemed to outperform the UM approach in terms of human interpretability and predictability, after the training period the average number of updates decreased which indicated that the proposed routes were matching the observed routes. There is however no evidence that indicated that the decision tree models are superior to MNL models in terms of predictability, the predictive accuracy of the decision tree models does show a consistent trend over the three tests while the traditional models demonstrated variable predictive accuracy depending on the data used.
Directions for future work
From the results, as described above, a number of directions for future works became apparent with which route guidance could be applied in real life situations. Firstly it was advised to develop a user interface to provide routing information and collect user feedback. Furthermore additional attributes, which were excluded due to the lack of information in the simulation study, about the transport network and the surrounding area should be considered. Also, it was advised to device a method to deal with qualitative attributes such as aesthetics (e.g. distance through natural environments) or driving comfort (number of stops and turns, which also affects the route choice of drivers. Lastly additional process of incorporating user preferences in route generation should be considered, for example by weighting certain attributes related to the users preferences in the process of computing generalized link costs.
102 | P a g e