The unfalsified control concept and learning

The "unfalsified control" concept is introduced as a framework for determining control laws whose ability to meet given performance specifications is at least not invalidated (i.e., not falsified) by the experimental data. The concept provides a clear perspective on the nature of learning in a deterministic setting. The approach is "model-free" in the sense that no plant model is required-only plant input-output data. When implemented in real time, the result is an adaptive robust controller which modifies itself whenever a new piece of data invalidates the present controller. A simple design example based on fixed-order LTI controllers and an L/sub 2/-inequality performance criterion is presented.<
>