The main message of this chapter is that constructing an enumerable architecture model and exploring the corresponding tradespace of architectures can help system architects to organize architectural decisions and to understand the coupling between them and their sensitivity for the metrics of interest.

We provided a procedure to analyze an architectural tradespace. The first step is to look at the fuzzy Pareto frontier—that is, the architectures that represent the best tradeoffs between the metrics of interest. Looking at common features among architectures on the fuzzy Pareto front can immediately reveal dominant decisions. Moreover, the Pareto front is a good place to start looking for good architectures to study in more detail, even though the condition of non-dominance cannot by itself serve as a basis for down-selecting to a small set of architectures.

Further analysis of the structure of the tradespace can reveal additional information. For example, the presence of clusters or stratification in the tradespace may indicate that the tradespace is dominated by a small set of driving decisions. Finding the set of driving decisions is a by-product of the tradespace mental model that can sometimes be even more important than finding a set of good architectures.

It is important to conduct sensitivity analysis on the results generated by an architectural ­model, especially when the uncertainty in the model parameters is large. In particular, sensitivity analysis can reveal architectures that are the most robust in responding to variations in the environment. Robust architectures are preferable because they are more likely to sustainably ensure value delivery to stakeholders over the system’s lifecycle.

Finally, the system architect can produce a hierarchy of architectural decisions that suggests the sequence in which decisions should be made, based on the four-quadrants method. Decisions in the first quadrant are highly sensitive and highly coupled, so they should be made first. Decisions in the fourth quadrant are relatively insensitive and uncoupled, so they can be relegated to the bottom of the tree. The mid-levels of the tree are filled out by grouping highly coupled decisions in a single trade study and having trade studies that are weakly coupled run in parallel.