Error Tolerant Plan Recognition: An Empirical Investigation

Few plan recognition algorithms are designed to tolerate input errors. We describe a case-based plan recognition algorithm (SET-PR) that is robust to two input error types: missing and noisy actions. We extend our earlier work on SET-PR with more extensive evaluations by testing the utility of its novel action-sequence representation for plans and also investigate other design decisions (e.g., choice of similarity metric). We found that SET-PR outperformed a baseline algorithm for its ability to tolerate input errors, and that storing and leveraging state information in its plan representation substantially increases its performance.