Application-level robustness and redundancy in linear systems

The paper quantifies the degradation in performance of a linear model induced by perturbations affecting its identified parameters. We extend sensitivity analyses available in the literature, by considering a generalization-based figure of merit instead of the inaccurate training one. Effective off-line techniques reducing the impact of perturbations on generalization performance are introduced to improve the robustness of the model. It is shown that further robustness can be achieved by optimally redistributing the information content of the given model over topologically more complex linear models of neural network type. Despite the additional robustness achievable, it is shown that the price we have to pay might be too high and the additional resources would be better used to implement a n-ary modular redundancy scheme