(posted on my blog: davenichols.net)

I've been interested in complex adaptive systems (cas) and genetic algorithms for many months now, and have a few things that I would love to model and test in some sort of stable framework. What I was missing was the specific nature of building adaptive agents and coordinating their interactions in a meaningful way. John Holland, the 'father of genetic algorithms', wrote this outstanding treatment on cas in 1995 and uses it to describe the work he and his colleagues undertook to model these systems in a framework called ECHO.

The first chapter lays out Holland's ideas about what these systems need in order to be modelled in this way. Namely, there are seven processes and properties which must be present (aggregation, tagging, nonlinearity, flows, diversity, internal models, and building blocks), and with these items, a proper cas model can be built and studied.

Next, Holland explains the nature of agents, each an instantiation of specific i/o, rules, and adaptive behavior. From there, the complex emergent properties and behaviors are explored, including the development of the model known as ECHO. Here, the meat of the book includes Holland's explanation of how to derive complex, emergent behavior from small sets of coded instructions which can be exchanged between agents and mutated. Finally, simulations of the modelling system and further advances toward a proper theory of cas are discussed.

This was the foundational work I was looking for to cement my non-professional understanding of cas modelling and how I might use my programming knowledge to build simple tests of emergent behavior.

I do want to specifically disagree with the previous reviewer(s?) who noted a couple of things. First, while they noted this book is not specifically intended for any professional, such critique misses the point of this book. Holland was working through a difficult modelling project during the mid-1990s, a time when cas was just beginning to be understood in terms of widespread computing resource availability. The work of Holland and his Santa Fe Institute collegues was groundbreaking, and Hidden Order is an excellent place to start for any non-professional interested in understanding how one might go about modelling what is often considered too complicated to model.

Second, Holland's previous work, Adaptation in Natural and Artificial Systems is considered a groudbreaking work in the field, and Holland himself is considered to be one of the most influential thinkers dealing with cas. While I have neither read Adaptation nor worked with cas professionally, this work is not intended for the 'layperson' (one review said it was 'too focused on computer modelling for a layperson'), but for someone, such as myself, who is a non-professional interested in computer modelling, mathematics, and emergent behavior. Not only was the entire book comprehensible, it was a perfect balance of technical discussion and broad explanation. Few books on such a difficult subject pull off this delicate balance, but Holland did so here.

Third, a critique in an earlier review utterly misrepresents Holland's treatment of economics and the Prisoner's Dilemma in this book. PD is used to show how a specific test of emergent behavior can be worked into a computer model of such behavior, it was not used to explore the views of economists on PD. Holland does not imply that only after artificial agents played PD were some results understood, he merely maintains that artificial agents realize many of the same stable strategies realized by human participants.

Holland's treatment of economics is very high-level and is never, ever meant to create a full view of any classical economic model. He offers a general view of some points of classical economics which can be useful in studying cas but never states that his explanation is even a small part of all economic thought and research. I think the reviewer who critiques Holland for 'suggest(ing) that economists are mostly ignorant of this strategy, when in fact it's widely used in economics' misses the point that this book was written a decade and a half before s/he wrote the review of the book. Economic modelling, like all modelling of the last fifteen years, has changed dramatically with the birth of cheap available computing resources.

That said, for other non-professional readers interested in complex adaptive systems, mathematics, computer modelling, and emergent behavior, this is an excellent place to start your study. You do need to be comfortable with logic and some basic modelling behavior, but it is hard to imagine that a reader is interested in the aforementioned subjects without being capable of handling this book. Four stars. ( )

Too focused on computer modeling for a layperson and not technical enough for a professional in any field involving the analysis of dynamic systems. This book lacks a well-chosen target audience.

Side note: According to the book's cover, Douglas Hofstadter called it "a monumental achievement," which calls into question Hofstadter's sanity. Did he actually read it? ( )

Not specific enough in any one discipline to be interesting to professionals in that discipline, and not general enough to add new insights for dynamic systems generally. I speak as an economist who knows enough biology (I think) to make these assessments.

I was disappointed to find some basic errors on economics. Holland says economists thought the mutual defection equilibrium of the Prisoners’ Dilemma game was the last word on that game, and implies the idea of cooperation in repeated play never occurred to anyone until artificial agents played the game repeatedly. Tripe. Game theorists have long known that mutual cooperation is an equilibrium for some versions of the repeated Prisoners' Dilemma; see any game theory textbook (even one for undergrads). Indeed, this has long been the canonical example used to illustrate the Folk Theorem.

I was also annoyed by Holland’s oversimplifications of classical economics. He then proceeds to report some interesting simulation work on financial markets, but his exposition suggests (1) that economists in general aren’t aware of this research strategy, and (2) that theoretical modeling has nothing to say about asset bubbles and crashes, etc. Actually, there is plenty of learning theory on financial markets, including bubbles and crashes. Recent work on large deviation theory and escape routes is applicable here. And in a sense, many E-unstable models are models of asset bubbles. The cobweb model with E-instability could be interpreted as a model of a bubbly asset market, if there are production lags on the supply side. As in, say, the HOUSING MARKET, for example.

None of this is to suggest that simulations aren’t a fruitful approach in economics. They're quite useful. But it is irksome that Holland suggests that economists are mostly ignorant of this strategy, when in fact it's widely used in economics.

As to the cross-disciplinary ambitions embodied in this book, I don’t think it does very well. In general, I think the differences in the systems Holland wants to include defy the kind of generality he has in mind without resorting to dynamic systems mathematics. Also, modeling human behavior is very different from modeling molecule, animal, or ecosystem behavior because humans are not only tatonnement learners, they are also purposeful, forward-thinking learners. Any insights gleaned from the study of merely locally adaptive systems are bound to fall short. Most humans aren’t Einstein, but some provision must be made for intelligence!

So: If one wants generality, one should simply study the relevant mathematics of dynamic systems, differential and difference equations. If one wants specifics for a given discipline, one should read in that discipline. If one wants cross-disciplinary insights that aren't mathematical, appropriate doses of Hayek might be instructive. ( )