Insights from Machine Learning for Plan Recognition

This paper explores the bene ts of adapting techniques from inductive concept learning to plan recognition. A powerful notion in concept learning is characterizing inductive systems by their bias, i.e. the implicit assumptions which justify the conclusions an inductive system produces. We present a spectrum of possible biases for plan recognition. We evaluate these biases based on how accurately they predict how people achieve goals in Unix. We also adapt algorithms for maintaining version spaces to produce a goal recognizer that runs in time sublinear in the number of potential goals. We show a factor of 5 to 10 speedup, on data collected in Unix, over a more straightforward approach which enumerates every potential goal.