Route recommendation for automotive domains is a knowledge-rich problem where the criteria for making decisions and the relative weight of these criteria can be personalized. The Adaptive Route Advisor serves as an intermediary between the driver and the complexity of the digital map. The agent and the driver interact to generate multiple route options, giving the driver a more satisfactory route than he or she would receive from a single-option route planner, and providing feedback from the driver that reflects his or her route preferences. The agent encodes these preferences in a user model that it uses to predict which route a driver will find most appealing.
Although interaction is in the driver's best interest if he or she wants a satisfactory route, the Advisor does not require it, and the amount of interaction is controlled by the driver. Ideally, as the agent better approximates the driver's cost function, interaction becomes less necessary and the agent becomes more autonomous. This low interaction requirement is crucial for in-car decision making where the driver's attention is necessarily focused elsewhere.
In general, our approach to developing advisory agents is to automatically and unobtrusively acquire value judgments by observing the user's actions in a domain, and to utilize interaction as an additional source of value judgments. The advisor generates a solution using its current user model, receives feedback from the user if its model is inaccurate, and corrects its model in areas relevant to the problem being solved.