part Two...

"The universality of the reactive paradigm is not very different from the universality of the epicycles. Both modes of universality ultimately arise from the mathematical fact that any function can be approximated arbitrarily closely by functions canonically constructed out of a suitably chosen 'basis set' whose members have a special form. [snip; mathematical illustrations] From this it follows that if any kind of system behavior can be described in functional terms [in the mathematical sense, not the biological sense, of function] it can also be generated by a suitably constructed combination of systems which generate the elements of a basis set, and this entirely within a reactive mode. But it is clear that there is nothing unique about a system so constructed; we can do the same with any basis set. All these systems are different from one another, and may be likewise different from the initial system whose behavior we wanted to describe. It is in this sense that we can only speak of simulation, and not of explanation, of our system's behavior in these terms.

Indeed, it is clear that if we are confronted with a system which contains a predictive model, and which uses the predictions of that model to generate its behavior, we cannot claim to understand the behavior unless the model itself is taken into account. Moreover, if we wish to construct such a system, we cannot do so entirely within the framework appropriate to the synthesis of purely reactive systems.

On these grounds, I was thus led to the conclusion that an entirely new approach was needed, in which the capability for anticipatory behavior was present from the outset. Such an approach would necessarily include, as its most important component, a comprehensive theory of models and of modeling. The purpose of the present volume in fact, is to develop the principles of such an approach, and to describe its relation to other realms of mathematical and scientific investigation. With these and similar considerations in mind, I proceeded to prepare a number of working papers on anticipatory behavior, and the relation of this kind of behavior to the formulation and implementation of policy. Some of these papers were later published in the "International Journal of General Systems". The first one I prepared was entitled, "Planning, management, policies, and strategies: Four fuzzy concepts", and it already contained the seeds of the entire approach I developed to deal with these matters. For this reason, and to indicate the context in which I was working at the Center, it may be helpful to cite some of the original material directly. The introductory section began as follows:

"It is fair to say that the mood of those concerned with the problems of contemporary society is apocalyptic. It is widely felt that our social structure is in the midst of crises, certainly serious, and perhaps ultimate. It is further widely felt that the social crises we perceive have arisen primarily because of the anarchic, laissez-faire attitude taken in the past towards science, technology, economics, and politics. The viewpoint of most of those who have written on these subjects revolves around the theme that if we allow these anarchies to continue we are lost; indeed, on way to make a name nowadays is to prove, preferably with computer models, that an extrapolation of present practices will lead to imminent cataclysm. The alternative to anarchy is management; and management implies un turn the systematic implementation of specific plans, programs, policies and strategies. Thus it is no wonder that the circle of ideas centering around the concept of planning plays a dominant role in current thought.

However it seems that the net effect of the current emphasis on planning has been simply to shift the anarchy we perceive in our social processes into our ideas about the management of these processes. If we consider, for example, the area of "economic development" of the underdeveloped countries (a topic which has been extensively considered by many august bodies), we find (a) that there is no clear idea of what constitutes "development"; (b) that the various definitions employed by those concerned with development are incompatible and contradictory; (c) that even among those who happen to share the same views as to the ands of development, there are similarly incompatible and contradictory views as to the means whereby the end can be attained. Yet in the name of developmental planning, an enormous amount of time, ink, money, and even blood is in the process of being spilled. Surely no remedy can be expected if the cure and the disease are indistinguishable.

If it is the case that planning is as anarchic as the social developments it is intended to control, then we must ask whether there is, in some sense, a "plan for planning" or whether we face an infinite and futile anarchic regress. It may seem at firs sight that by putting a question in this form we gain nothing. However, what we shall attempt to argue in the present paper is that, in fact, this kind of question is "well-posed" in a scientific sense: that it can be investigated in a rigorous fashion and its consequences explored. Moreover, we would like to argue that, in the process of investigating this question, some useful and potentially applicable insights into planning itself are obtainable.

The notion of how planning could go wrong was of course of primary interest to the Center; indeed, for months I had heard a succession of discouraging papers dealing with little else. It seemed to me that by elaborating on this theme I could establish a direct contact between my ruminations and the Center's preoccupations. My preliminary discussion of these matters ended as follows:

We would like to conjecture further that, for any specific planning situation, each of the ways in which planning can go wrong will lead to a particular kind of syndrome in the total system (just as the defect of any part of a sensory mechanism in an organism leads to a particular array of symptoms). It should therefore be possible, in principle, to develop a definite diagnostic procedure to "trouble-shoot" a system of this kind, by mimicking the procedures used in neurology and psychology. Indeed, it is amusing to think that such planning systems are capable of exhibiting syndromes (e.g. of "neuroses") very much like and indeed analogous to those manifested by individual organisms.

Such considerations as these led naturally to the general problems connected with system error, malfunction or breakdown, which have always been hard to formulate, and are still poorly understood. Closest to the surface in this direction, especially in the human realm, were breakdowns arising from the incorporation of incorrect elements into the diagram shown above.... faulty models, inappropriate choice of effectors, etc. I soon realized, however, that there was a more profound aspect of system breakdown, arising from the basic nature of the modeling process itself, and from the character of the system interactions required in the very act of imposing controls. These were initially considered under the heading of "side effects", borrowing a medical terminology describing unavoidable and usually unfortunate consequences of employing therapeutic agents (an area which of course represents yet another branch of control therapy). As I used the term, I meant it to connote unplanned and unforeseeable consequences on system behavior arising from the implementation of controls designed to accomplish other purposes; or, in a related context, the appearance of unpredicted behavior in a system build in accordance with a particular plan or blueprint. Thus the question was posed: are such side effects a necessary consequence of all control? Or is there room for hope that, with sufficient cleverness, the ideal of the "magic bullet", the miraculous cure which specifically restores health with no other effect, can actually be attained?

Since this notion of side effects is so important, let us consider some examples. Of the medical realm we need not speak extensively, except to note that almost every therapeutic agent, as well as most diagnostic agents, create them/ sometimes spectacularly so, as in the thalidomide scandal of some years past. We are also familiar with ecological examples, in which man has unwittingly upset "the balance of nature" through injudicious introduction or elimination of species in a particular habitat; well known instances of this are the introduction of rabbits to Australia, to give the gentlemen farmers something to hunt on the weekend; or the importation of the mongoose into Santo Domingo, in the belief that because the mongoose kills cobras it would also eliminate local poisonous snakes such as the fer-de-lance. Examples from technology also abound; for instance we may cite the presence of unsuspected oscillatory modes in the Tacoma Bay Bridge, which ultimately caused it to collapse in a high wind; or the Ohio Turnpike, which was built without curves on the theory that curves are where accidents occur; this led to the discovery of road hypnosis. Norbert Weiner warned darkly of the possibility of similar disastrous side effects in connection with the perils of relying on computers to implement policy. He analogized this situation to an invocation of magical aids as related in innumerable legends and folk tales; specifically, such stories as "The Sorcerer's Apprentice", "The Mill Which Ground Nothing but Salt", "The Monkey's Paw", and "The Midas Touch". And of course, many of the social and economic panaceas introduced in the past decades have not only generated such unfortunate side effects, but have in the long run served to exacerbate the very problems they were intended to control.

There is, however, a class of planning difficulties which do not arise from such obvious considerations and which merit a fuller discussion. This class of difficulties has to do with the problem of side effects; as we shall see, these will generally arise, even if the models system is perfect and the effectors perfectly designed and programmed, because of inherent system-theoretic properties. Let us see how this comes about.

We stressed in that paper how the fact that all of the state variables defining any particular system S are more or less strongly linked to one another via the equations of motion of the system, taken together with the fact that the many state variables not involved in a particular functional activity were free to interact with other systems in a non-functional or dysfunctional way, implied that any particular functional activity tends to be modified or lost over timer. This, we feel, is a most important result, which bears directly on the "planning" process under discussion. The easiest way to see this is to draw another corollary from the fundamental proposition that only a few degrees of freedom of a system S are involved in any particular functional activity of S.

The corollary (4) is true even of the best models, and it is this corollary which bears most directly on the problem of side effects. Let us recall that S is by hypothesis a real system, whereas M is only a model of a particular functional activity of S. There are thus many degrees of freedom of S which are not modeled in M. Even if M is a good model, then, the capability for dealing with the non-functional degrees of freedom in S have necessarily been abstracted away. And these degrees of freedom, which continue to exist in S, are generally linked to the degrees of freedom of S which are modeled in M, through the overall equations of motion which govern S.

Now the planning process requires us to construct a real system E, which is to interact with S through a particular subset of the degrees of freedom of S (indeed, though a subset of those degrees of freedom of S which are modeled in M). But from our general proposition, only a few of the degrees of freedom of E can be involved in this interaction. Thus both E and S have in general many "non-functional" degrees of freedom, through which other, non-modeled interactions can take place. Because of the linkage of all observables, the actual interaction between E and S specified in the planning process will in general be affected. Therefore, we find that the two following propositions are generally true: (a) An effector system E will in general have other effects on an object system S than those which are planned; (b) The planned modes of interaction between E and S will be modified by these effects. Both of these propositions describe the kind of thing we usually refer to as side effects. As we see, such side effects are unavoidable consequences of the general properties of systems and their interaction. They are by nature unpredictable, and are inherent in the planning process no matter how well that process is technically carried out. As we pointed out in our previous paper, there are a number of ways around this kind of difficulty, which we have partially characterized, but they are only applicable in special circumstances.

The basic principle struggling to emerge here is the following: The ultimate seat of the side effects arising in anticipatory control, and indeed of the entire concept of error or malfunction in system theory as a whole, rests on the discrepancy between the behavior actually exhibited by a natural system, and the corresponding behavior predicted on the basis of a model of that system. For a model is necessarily an abstraction, in that degrees of freedom which are present in the system are absent in the model. In physical terms, the system is open to interactions through these degrees of freedom, while the model is necessarily closed to such interactions; the discrepancy between system behavior and model behavior is thus a manifestation of the difference between a closed system and an open one. This is one of the basic themes which we shall develop in detail in the subsequent chapters.

My initial paper on anticipatory systems concluded with several observations, which I hoped would be suggestive to my audience. The first was the following: that it was unlikely that side effects could be removed by simply augmenting the underlying model, or by attempting to control each side effect separately as it appeared. The reason for this is that both of these strategies face an incipient infinite regress, similar to that pointed out by Gödel in his demonstration of the existence of unprovable propositions within any consistent and sufficiently rich system of axioms. Oddly enough, the possibility of avoiding this infinite regress was not entirely foreclosed; this followed in a surprising way from some of my earliest work on in relational biology, which was mentioned earlier:

There are many ramifications of the class of systems developed above, for the purpose of studying the planning process, which deserve somewhat fuller consideration than we have allowed. In this section we shall consider two of them: (a) how can we update and improve the model system M, and the effector system E, on the basis of information about the behavior of S itself and (b) how can we avoid a number of apparent infinite regresses which seem to be inherent in the planning process?

These two apparently separate questions are actually forms of the same question. We can see this as follows. If we are going to improve, say, the model system M, then we must do so by means of a set of effectors E' . These effectors E' must be controlled by information pertaining to the effect of M on S; i.e. by a model system M' of the system (S+M+E). in other words, we must construct for the purpose of updating and improving M a system which looks exactly like it except that we replace M by M', E by E', and S by S+M+E. But then we may ask how we can update M' ; in this way we see an incipient infinite regress.

There is another infinite regress inherent in the discussion given of side effects in the preceding section. We have seen that the interaction of the effectors E with the object system S typically give rise to effects in S unpredictable in principle from the model system M. However, these effects too, by the basic principle that only a few degrees of freedom of S and E are utilized in such interactions, are capable of being modeled. That is, we can in principle construct a new model system M1 of the interaction between S and E, which describes interactions not describable in M. If these interactions are unfavorable, we can construct a new set of effectors, say E1, which will steer the system S away from those side effects. But just as with E, the system S will typically interact with E1 in ways which are in principle not comprehensible within the models M or M1; these will require another model M2 and corresponding new effectors E2. In this way we see another incipient infinite regress forming. Indeed, this last infinite regress is highly reminiscent of the "technological imperative" which we were warned against by Ellul and many others. Thus the question arises; can such infinite regresses be avoided?

These kinds of questions are well-posed, and can be investigated in system theoretic terms. We have considered questions like these in a very different connection; namely, under what circumstances is it possible to add a new functional activity to a biological organization like a cell? It turns out that one cannot simply add an arbitrary function and still preserve the organization; we must typically keep adding functions without limit. But under certain circumstances, the process does indeed terminate; the new function is included (though not just the new function in general) and the overall organization is manifested in the enlarged system. On the basis of these considerations, I would conjecture that (a) it is possible in principle to avoid the infinite regresses, and in particular to find ways of updating the model M and the effectors E' (b) not every way of initiating and implementing a planning process allows us to avoid the infinite regress. The first conjecture is optimistic; there are ways of avoiding this form of the "technological imperative". The second can be quite pessimistic in reference to our actual society. For if we have in fact embarked on a path for which the infinite regresses cannot be avoided, then we are in serious trouble. Avoiding the infinite regresses means that the developmental processes will stop, and that a stable steady-state condition can be reached. Once embarked on a path for which the infinite regresses cannot be avoided, no stable steady-state condition is possible. I do not know which is the case in our own present circumstances, but it should at least be possible to find out.

The theoretical principle underlying this analysis of failure in anticipatory control systems is not wholly negative. In fact, we shall argue later that it also underlies the phenomena of emergence which characterize evolutionary and developmental processes in biology. It may be helpful to cite one more excerpt of a paper originally prepared for the Center Dialog, which dealt with this aspect:

I will take a look for something along the lines of John M's request, although those who have a copy of this book are invited to do the same. Anticipatory Systems was and still is a revolutionary little book, which was never really understood enough, in my view, to make it as infamous as it would otherwise have been in science. I used to worry that my father would become the target of someone like the Unabomber, a Luddite who hated science and technology and was willing to kill to stop forward progress. I think the internet may have accomplished that, if it had existed to the degree it does now, in the 1970's and 1980's.