Symposium on Complex Systems Engineering/Schedule/Thursday 1c comments
From CSWiki
M Kuras:
Tyson Browning’s paper may well be one of the most important at this symposium. It is one of the few that truly steps outside the bounds of the accepted conventional wisdom on the nature of (developmental) Programs and their outcomes.
He correctly notes that Programs cannot be hierarchically decomposed. He borrows from chemistry to convey this notion: the idea of a “capability molecule” in which five systems are the analogs of “atoms” that “bond” to form the “molecule” of the Program.
In my own vocabulary (see my papers) what Tyson calls systems I refer to as “regimes.” And I would continue to refer to a Program as a system (a complex-system, to be exact). This can of course cause confusion.
He singles out five systems (my regimes) in a Program: the product, the process, the tools, the goals, and the organizational ones. That is because he is (correctly) focused on development, but this leads to the omission of a sixth regime that is crucial in understanding what happens in the development of things like (say) modern air craft, manned and unmanned (and more importantly fleets of such aircraft). This is the operational regime. Development does not stop once initial versions of products are used; and in fact that usage strongly shapes future development.
There are additional regimes as well (what he would call systems) that help to comprise or that strongly influence his capability molecule of a Program.
There is much to be considered about his fundamental notion that Programs cannot be conventionally decomposed. The Program is an “emergent” consequence of the contributions of many constituent systems (in his vocabulary). I truly hope this happens.
He correctly notes that there are many relationships among his five systems (I would say across the regimes) that comprise his capability molecule of a Program. My second comment hinges on this notion of a “relationship.”
Implicit in Tyson’s treatment of DSMs (and derived DMMs) and elsewhere is the notion of networks to represent or to help model the relationships that he wishes to highlight. It seems to me that Tyson is relying (implicitly) on Metcalf networks (in fact canonical Metcalf networks) as his fundamental conceptual notion for doing this. He doesn’t have to do that and in fact, a more generalized notion of networks will greatly aid his efforts.
As we all know, there are Sarnoff and Metcalf networks. There are also Reed (or combinatoric) networks. I really believe that if Tyson uses Reed networks he will greatly accelerate his research efforts. (For a variety of reasons all of the glamour and attention is attached to Metcalf networks right now.)
The following is an excerpt from my paper on a Multi Scale Definition of a System released by MITRE last year…
“For those so inclined, an entire RELATIONSHIP can be considered to be a sparse Reed (or combinatoric) network with P-VALUES as its nodes, and expressions as its links. The links in a Reed network are not directional (either unidirectional or bidirectional) as they usually are in both Sarnoff and Metcalf networks. Links in Sarnoff and Metcalf networks are usually used to model flows; and directionality is essential. Links in Reed networks are non-directional or associational. Many nodes can be associated by a single link (and many links can involve a single node). Almost all functions in mathematics are special cases of RELATIONSHIPS.”
This excerpt is from a discussion of relationships. What is important to appreciate here is that Reed network links are not directional. No “flows” are necessarily involved (although that is a possibility when the Reed network is also a Metcalf network).
When I say canonical, I mean, for example that in a Metcalf network there can be only one link between nodes, and nodes are not linked to themselves. But I digress.
I would also encourage Tyson to examine closely the processes of evolution as we have come to understand them. These apply to his Part II discussion, especially if he accepts (as I do) that learning and evolution are essentially synonyms for the same phenomenon considered as distinct scales. I think that Doug Norman is keen on pressing this particular focus.
Roger Burkhart's paper also deserves immediate comment:
Roger makes some important observations:
“Large numbers of elements that interact in decentralized ways, giving rise to emergence across multiple levels of causation, make a complex system difficult to understand and analyze, much less engineer, using traditional techniques of closed-form analysis or prediction.”
“Just because a system is complex, however, does not mean that it is incapable of being engineered, or that traditional methods of systems engineering management are not required to define and trace the realization of a system and its relationships to stakeholder needs and requirements.”
However, Roger also points to UML and other derivatives of first order logic as, perhaps, the key enabler.
Modeling and simulating a complex-system (I do not mean here a SoS) requires that we understand the fundamental distinctions between complex-systems and “ordinary” systems.
We also need to extend M&S techniques to SoS, most of which are closely related to ordinary systems.
Complex-systems function and have form (both substance and structure) that are available only at multiple scales of human conceptualization. Ordinary systems (their functionality and form) and many SoS can be adequately understood at a single scale of conceptualization. Models and simulations of complex-systems need to take the need for multiple scales of conceptualization into account. In particular, such models and simulations (in order to be useful) need to confront the dependencies that clearly exist across multiple scales of conceptualization. (I like Russ Abbott's introduction of the word "entailment" in this regard) The fundamental obstacle is that first order logic and its derivatives (although appropriate for modeling and simulating ordinary systems and even SoS) do not work across multiple scales of conceptualization.
Moreover, not only is it important to gauge the influence of functionality and form at “lower” scales of conceptualization on those at “higher” scales of conceptualization, so is the reverse: how functionality and form at higher scales influences functionality and form at lower scales.
So we really have two challenges. We need to extend modeling and simulation (as an aid to engineering) to SoS and to complex-systems.
Many complex-systems, and even some ordinary systems and SoS, are also social systems (behavior or functionality as well as form are best expressed in terms of people or autonomous agents). I think that Roger would agree that ABMs are the best tools that we have at this time for such systems. To the extent that such systems evolve (or, equivalently, learn) and such developments are important to the problem at hand, the processes of evolution are important to incorporate into our models and simulations. This requires a deeper appreciation of the processes of evolution by those engaged in developing models and simulations. I would encourage you to speak to Doug Norman about this.
Dan and Yaneer illustrate why so much attention (and glamour) are currently attached to the use of canonical Metcalf networks as an analytical tool! Their observations and analysis are valuable additions to our body of knowledge. There are important implications arising out of their observations.
That said, I would encourage both of them to consider the use of more than just the canonical Metcalf network to capture and distill “real world” data so that we can better comprehend the patterns that the “real world” exhibits.
When we use just one of the available models, there is the risk that we are forcing some of the "real world" data into pattern templates to which they do not really belong.
To quickly review, there are (at least) Sarnoff, Metcalf and Reed networks. All of them also have canonical forms.
Sarnoff networks have TWO sorts of nodes: sources and sinks; and ONE sort of link, unidirectional. In a canonical Sarnoff network, there is just one source node and any number of sink nodes, all of which are connected to the source node by one link.
Metcalf networks have one sort of node and several sorts of links. In a canonical Metcalf network, the links are bi-directional (or “undirected” as some say), and there is at most one link between any pair of nodes with the stipulation that no node is connected to itself.
Reed networks are also termed combinatoric networks. They have one sort of node, and the links are associational. (The use of the term “undirected” for some Metcalf networks can cause confusion here.) In a Reed network, any number of nodes can be connected by one link (and any number of links can involve one node).
Sarnoff and Metcalf networks are almost always used to capture flows.
The “capacity” of a canonical Sarnoff network (expressed as a number of links) is proportional to the number of nodes. The capacity of a canonical Metcalf network is proportional to the square of the number of nodes. The capacity of a Reed network is proportional to 2 raised to the nth power, where n is the number of nodes.
To reiterate, I believe that it is important to consider fitting “real world” data into more than just one of the available templates (canonical and non-canonical Sarnoff, Metcalf and Reed network models).
