Model identification and control from a probabilistic viewpoint

This paper considers a probabilistic approach to modeling and control, where probabilities are conditional, based upon input-output data. For this approach, we answer the following questions. Are scalar measures of identification quality good predictors of probability of performance? And, does improving a scalar measure of robustness necessarily imply improving the probability of performance? It turns out that the answer to both of these questions is no. We analyze the underlying mechanisms for such phenomena and provide general propositions to predict their occurrence. A missile autopilot example is used throughout the paper to illustrate the results