Reference: Gans, A. & Hayes-Roth, B. NEWWATCH: Learning Interrupted Strategies by Observing Actions. Knowledge Systems Laboratory, March, 1990.
Abstract: An expect can normally demonstrate a strategy by solving individual problems. However, it is more difficult to articulate the set of general rules used to select actions. The WATCH project was designed to "watch over the shoulder" of an expect, to automatically learn strategic knowledge by observing the sequence of actions that were performed. The original WATCH program was limited by the assumption that the entire sequence of problem-solving actions could be induced into a single hierarchical strategy tree. Any interruptions in this single strategy, whether a simple opportunistic sequence or a second parallel strategy, were simply treated as noise that confused the system. The NEWWATCH extension has been implemented to recognize various kinds of interruptions, thereby greatly increasing the range of strategies that the WATCH system can learn.