Automated Planning and Intent Recognition

Automated planning is the process of determining a sequence of actions that will achieve a desired result. The planner reasons about the differences between the current state of the world and the desired goal state. The problem is made more challenging when the world state is not fully known, when time is a factor, when action effects are probabilistic, and so forth. SIFT researchers have led efforts in planning under real-time constraints (CIRCA), planning as a hierarchical decomposition exercise (SHOP3), and planning under model uncertainty.

Automatic intent recognition is the dual of planning - the system attempts to derive the user's plan from observation of actions. Also referred to as plan recognition, it is a critical challenge for intelligent user interfaces (to determine what a user is trying to do, and how the UI can offer help), computer security (to determine the objectives of an attacker), sketch understanding, natural language understanding, and other contemporary problems. SIFT researchers pioneered the use of Bayesian probabilistic methods in intent recognition. In cooperation with university researchers, SIFT has developed the Yappr system for intent recognition that achieves new levels of computational efficiency by using sophisticated techniques based on parsing.