The four kinds of ACS models maintain the four kinds of stable representations needed along the course of system flow. But each is only the ``top'' of series of representations of different aspects of the world derived from sensors and other inputs. There may be several intermediate representations of state vector values in between sources and best possible estimates. Also, many submodels of primary models are used only for purposes of estimating aggregate values in others, so mainly serve as intermediate estimators.
Whenever it is necessary to maintain intermediate representations, Estimator components should be constructed. If they take the same form as the main models themselves, architectures assume an extensible, recursive pattern in which outputs of one stage become inputs for the next. This also permits code-sharing of implementations of common features of estimators and top-most models.
These considerations may also be applied to ``mandatory'' models. Sometimes even these don't really need stable representations. In this case, the steps may be applied in reverse: State vectors may be passed around through filters as data values without ever ``resting'' in a model.
Determine the numbers and kinds of estimator stages that are needed, by considering:
Each of these estimators can then be structured, connected, updated and controlled in the same fashion as top-most models.