| Computational models of motor control vary widely in method and application, often demanding extensive calculations in order to minimize cost functions and/or extensive implicit knowledge of physical properties of the entire system (e.g., Kawato et al, 1992, 1996). Although such models have had demonstrated success in predicting wide ranges of movement behaviour they are limited in their ability to capture the variability of performance normally displayed by the human system. Current models are similarly unable to account for more than a confined task type, often making restrictions such as limiting performance to open-loop control of simple reaching movements. Presented here are applications of hidden Markov models (HMMs)---a tool known in the machine learning literature for representing first-order dynamical systems in a stochastic fashion---and its generalization, dynamical Bayesian networks (DBN), to human movement. The HMM approach enables the representation of probabilistic relations between elements of a system for the expression of system dynamics the structure of DBNs is such that combinations of HMMs may be used to build a model of the target system based on biological knowledge, or hypothesis. An extensive discussion of methods for training and assessing both HMMs and DBN is presented within a motor control context. Further, four experiments are described which examine the utility and efficacy of these methods for the representation, recognition, and production of both discrete and continuous motor tasks. General comparisons are made throughout between DBN/HMM techniques and other modeling alternatives for representing time-varying biological movement signals. Extensions of the HMM/DBN framework are provided along with suggestions for future applications and possible implications to current theories of motor control. |
