Evolving Models for Signal Processing

Tuesday December 12, 2000, Morning

Modeling is a key element of signal processing, and perhaps the most difficult. Success requires having quality data, a proper space of mathematical models, and an appropriate search algorithm. Neural networks and other nonlinear models have proved useful in capturing many of the relevant dynamics of systems of interest. Optimizing these and even linear models by evolutionary algorithms can provide additional advantages. The researcher is free to search a wider domain of models and can in many cases avoid the two-step approach of hypothesizing a model and testing it for overfitting. Information criteria can be used directly to evaluate model in terms of goodness-of-fit and degrees of freedom, simultaneously. Some examples of the use of information statistics and other evolved models in signal processing domains will be discussed.