A family of multiple-model smoothers for signal classification
Phillip L. Ainsleigh, Tod Luginbuhl · 2005
In general, signal classification requires methods for constructing the classifier decision function from training data, as well as methods for evaluating the trained decision function for unlabeled data. When class assignments are made based on the time evolution of characteristic features, the classifier often employs a state-space tracking algorithm. And when the signal characteristics can change abruptly, multiple-model tracking algorithms are used. A pre-requisite for training such models is the ability to accurately estimate the time-varying probability distributions of the states and model assignments. This work examines a family of multiple-model smoothers, or forward-backward algorithms, that approximate the desired posterior distributions. Simulations are used to judge the tracking capabilities of the smoothers, which provide an indication of the estimation accuracy for the resulting distributions.