The state of the art in computer technology has advanced to the point where systems for generating driving directions between two points are commonplace. There are several web sites offering street-level driving directions, and several in-car systems available as an option on purchased or rented cars. The availability of vector road map digital representations (digital maps) and large-capacity, fast processors enable this technology, but little attention has been paid to ensuring that the interface is flexible enough to deliver satisfactory routes to users that have different preferences.
Current systems for route advice compute solutions using a shortest path algorithm to find the minimal-cost route from the origin to the destination. Some systems fix the cost as the estimated travel time, while others allow the user to choose between the shortest path, the quickest, or the ``most scenic'' one. In all cases, the system then describes the route to the user with little or no recourse if the driver finds the route unsatisfactory. These systems disregard the fact that driving occurs in a rich environment where many factors influence the desirability of a particular route. For example, some drivers may prefer the shortest route as long as it does not have too many turns, or the fastest route as long as it does not go on the highway. The relative importance of these factors varies among individuals, and drivers may not know themselves what they value most in routes.
In this paper, we describe the Adaptive Route Advisor, an adaptive user interface [7] agent that recommends routes from a source address on a road network to a destination address. Given a routing task, the Route Advisor interacts with the driver to generate a route that he or she finds satisfactory. Initially, the agent proposes a small number of possible routes, taking into account the driver's preferences if known. The driver can then request solutions that differ from the initial options along some dimension. For instance, the driver may request a route that is simpler, even if it means a longer route. The driver and the Route Advisor continue to work together, generating and evaluating different solutions, until the driver is satisfied. During this process, the agent unobtrusively collects information about the driver's choices and uses these data to refine its model of the driver's preferences for the next task.
If we define the level of autonomy as the number of interactions with a user, with zero interactions implying total autonomy, the user adjusts the autonomy of the Adaptive Route Advisor by continuing to request routes until satisfied. Since the system bases its initial route suggestion on the user model, the user's satisfaction with the initial route depends on the accuracy of the model. Ideally, after a reasonable number of interactions, the agent's user model will be accurate enough so that in most cases the driver will be satisfied with the routes proposed and that additional interactions will not be required. This is particularly important in a driving environment where the demand on the driver's attention must be limited.
The Adaptive Route Advisor is designed for in-car use. It is a Java application that functions as a resource-light network client, suitable for mobile environments with a wireless communication infrastructure. The remote servers provide resource-intensive functions such as routing and geolocation.1 Although the current version does not yet take advantage of information available from mobile deployment (primarily current and past locations from the Global Positioning System), and the interface is not fully optimized for limited input and output resources common in vehicles, we intend to more fully assimilate the Adaptive Route Advisor in a mobile environment in future work.
The pages that follow describe our approach in more detail and present the results of an experiment in personalizing the user model from rankings of routes generated from static preference models. First we present the overall system architecture, including the route generation component, the adaptation method the Route Advisor uses to personalize the user preference model, and the user interface that presents route options to the user and gathers preference feedback. We then report on an experiment adapting a preference model to human subjects and its results. Next, we present our approach to handling hidden attributes and outline planned improvements to the agent. Finally, we summarize the Adaptive Route Advisor and describe its relevance to more general problems.