Goal-Directed Knowledge Acquisition

Agents with incomplete knowledge of their actions can either plan around the incompleteness, ask questions of a domain expert, or learn through trial and error. We present and evaluate several approaches to formulating plans under incomplete information, using the plans to identify relevant (goaldirected) questions, and interleaving acting, planning, and question asking. We show that goal-directed knowledge acquisition leads to fewer questions and lower overall planning and re-planning time than naive approaches that ask many questions, or learn by trial and error. Moreover, we show that prioritizing questions based on plan failure diagnoses requires less questions on average.