Adaptive Goal Recognition

Because observing the same actions can warrant different conclusions depending on who executed the actions, a goal recognizer that works well on one person might not work well on another. Two problems that arise in providing user-specific recognition are how to consider the vast number of possible adaptations that might be made to the goal recognizer and how to evaluate a particular set of adaptations. For the first problem, we evaluate the use of hillclimbing to search the space of all combinations of an input set of adaptations. For the second problem, we present an algorithm that estimates the accuracy and coverage of a recognizer on a set of action sequences the individual has recently executed. We use these techniques to construct Adapt, a recognizer-independent unsupervised-learning algorithm for adapting a recognizer to a person's idiosyncratic behaviors. Our experiments in two domains show that applying Adapt to the BOCE recognizer can improve its performance by a factor of two to three.