Misfit versus latency (Q5949072)

A new framework for modeling dynamic linear time-invariant systems is developed. It combines ideas of the prediction error approach and of the behavioral approach. The set of all conceivable input/output time series compatible with the model is called the behavior of the system. A model is found by selecting the time series of the behavior which is closest to the observed data. A user-defined distance measure is used and gives the misfit of the resulting model. The authors develop the idea of obtaining models that `explain' the observed data by modifying the data (misfit) or by introducing new `unobserved' additional data (latency) or, by applying a combination of misfit and latency. The proposed model selection procedure can be rephrased as an optimization problem that only depends on the model parameters. Experiments illustrate the potential of this approach.