Nate Blaylock, James Allen
We present our work on using statistical, corpus-based machine learning techniques to simultaneously recognize an agent's current goal schemas at various levels of a hierarchical plan. Our recognizer is based on a novel type of graphical model, a Cascading Hidden Markov Model, which allows the algorithm to do exact inference and make predictions at each level of the hierarchy in time quadratic to the number of possible goal schemas. We also report results of our recognizer's performance on a plan corpus.
Subjects: 8. Enabling Technologies; 1.11 Planning