Melinda T. Gervasio - Wayne Iba - Pat Langley
Institute for the Study of Learning and Expertise
2164 Staunton Court, Palo Alto, California 94306
{gervasio,iba,langley}@isle.org
In this paper, we describe INCA, an adaptive, advisable assistant for crisis response. The system lets users guide the search toward particular schedules by giving high-level, operational advice about the solutions desired. Because traces of user interactions provide information regarding the user's preferences among schedules, INCA can draw on machine learning techniques to construct user models that reflect these preferences. We characterize the modeling task as that of learning a weight vector for a linear evaluation function that will lead to the same pairwise preferences between schedules as the user. INCA adapts to individual users by adjusting the weights on its evaluation function using a perceptron-type learning algorithm. To evaluate the system's adaptive capabilities, we designed an experiment involving four types of synthetic users that differed in their evaluation functions and in the level of advice they provide. We present experimental results showing that INCA achieves better performance with more specific advice and with learning, even on users with nonlinear evaluation functions or functions with additional, unknown parameters.