Falsification techniques

Abstract

Falsifiability of an identification problem was defined as the feasibility of determining whether a given model M is contradicted by data (c,d), that is, the feasibility of computing a set ℬ(c,d) of ‘unfalsified models’. That condition is obviously important in its own right for the feasibility of testing a final model. The falsification rule is then M ∉ℬ (c,d). But it is also the basis for designing a stopping rule for the refinement of the experiment (the inner loop) in the sequential identification procedure outlined in section 4.4. In that case one wants to test whether all models in a class ℳ are contradicted by data, and the falsification rule becomes ℳ ∩ ℬ (c,d) = ∅. Due to the randomness in the data the set ℬ(c,d) is not unique and not necessarily the same for model and structure. Falsification techniques deal with the problems of finding suitable sets ℬ(c,d) for the two cases.