Standard techniques (Yule-Walker) are available for learning Auto-Regressive process models of dynamical processes. When sensor noise means that dynamics are observed only approximately, learning has still been achieved via Expectation-Maximisation (EM) together with Kalman Filtering. This cannot handle more complex dynamics, involving multiple classes of motion. For that case, we show here how EM can be combined with the {\sc Condensation} algorithm, which is based on propagation of random sample-sets. Experiments have been performed with visually observed juggling, and plausible dynamical models are found to emerge from the learning process.

File

In  Advances in Neural Information Processing Systems,

Publisher  MIT Press

Details