Unsupervised Model-Based Prediction of User Actions

In this paper we derive algorithms that construct simple continuous-density hidden Markov models based on unlabeled observations of user actions. These models are then be used to predict future actions of the user. The algorithms were designed for domains where user-generated training data are costly to obtain and, consequently, the number of free parameters must be kept low. More specifically, we derive an algorithm that constructs -compact CDHMMs, that is Markov chains with an irreducible number of states, based on real-valued observations from users performing tasks. We show that the worstcase computational complexity of the algorithm is a second-order polynomial in the number and length of the tasks. We derive an optimal estimator for the expectation of future observations, based on recent observations, and give the computational complexity for the estimator. We also present a novel application: predictive robot programming. We show that the algorithms derived in this paper can recognize and predict the waypoints of robot programs in a rotationand translation-independent fashion, and we also derive the maximum-likelihood equations for locally optimal scale-invariant prediction. We present experimental results of the performance of our algorithms for predictive robot programming.