Setting Goals in Cybernetic Systems
Date Published
W. Ross Ashby
Borden1 Neurological Institute, England
Published in Cybernetics, Artificial Intelligence and Ecology: Proceedings of the Fourth Annual Symposium of the American Society for Cybernetics, ed. H. W. Robinson and D. E. Knight (Spartan Books, 1972), pp. 33–44. Reprinted in Mechanisms of Intelligence: Ashby's Writings on Cybernetics, ed. R. C. Conant (Intersystems, 1981), pp. 115–126. Scan of the reprint on ashby.info.
Getting clear the matter of goals is of the first importance in cybernetics, for most applications of cybernetics start with someone saying "I want . . . ". Here I am thinking of cybernetics not as a way of explaining things, but as a new science and technical method enabling us to tackle practical problems that would otherwise defeat us by their complexity. Coordinating the traffic around an airport, stabilizing the flows of money between international banks, normalizing the composition of the blood in a patient without kidneys—all these must start with the question "What do you want?" The process itself will end at the goal; the cybernetician's thoughts must start there.
What is a goal? We all know something about it, and a child of three frequently says "I want . . . ," so we all start with a personal awareness of intention, purpose, need, desire. But when we ask how a machine can have a desire, we find ourselves in difficulties. The difficulty becomes even greater if the system that is to have the goal is not even one machine but a mixture of machines and men, with the goal involving only the whole, not the parts. How can such a system have a purpose or desire?
The solution of this first difficulty, I suggest, is to do what the psychologists have done for a century: drop the introspectional aspect and turn to the aspect of behavior. Stop asking "Does this system feel a want?" and ask instead "How does it behave?"
Those whose knowledge of these matters is mainly introspectional may well hesitate to abandon their main source of information. But a century's work in psychology has shown that the introspectional approach, though vivid and apparently unquestionable, is in fact grossly unreliable. Look, for instance, at a piece of uncolored (white) paper: if anything is obvious and trustworthy, it is that there is no red present. Yet the physicists have convinced us that what we see is not the paper but a message from the retina saying that the three primary colors "balance." The introspective viewer sees only his own retina, not beyond it.
A report based on introspection is in fact simply the output of the brain's final, verbalizing stage. Such report can give only a coded version of what is happening earlier in the processes; to take a coding literally is an evident mistake. Psychoanalytic studies have shown in innumerable cases how mistaken a person can be when he describes his own motives or goals. Briefly, the introspectional approach, in science, has so far proved to be either just useless or positively misleading.
But if a goal is not a want, what is it? Ever since McDougall,1 psychologists have understood that it can be treated equivalently as a way of behaving. "Take a timid animal," he wrote, "such as a guinea-pig, from its hole or nest, and put it upon the grass plot. Instead of remaining at rest, it runs back to its hole; push it in any other direction, and, as soon as you withdraw your hand, it turns back towards its hole." Just the same behavior is characteristic of the missile that persists in going toward a source of infrared rays, that reasserts its direction if diverted, and that will change direction if the source moves.
The experiences of a century in psychology and of thirty years in automatic control systems have shown that, for practical purposes, we can achieve clarity by replacing the idea of a felt need with the idea of a focus in a stable dynamic system.
When the system is as simple, essentially, as a missile that "seeks" infrared rays, the thesis will probably not be disputed. But what of the more complex? What of natural evolution, for instance, with organisms apparently developing their own goals? What of man? Can he not choose his own goals? Cannot the cybernetician make a machine that can choose its own goal?
To get our ideas untangled, let us first take the case of natural evolution, since here the facts are, today, beyond dispute. This case is that of a planet, subject both to unchanging laws (such as those of gravity, optics, hydrodynamics) and, over all of 10¹⁰ years, to a constant energy input. Through all this time, photons of the visible wavelengths have poured in and have left at infrared lengths to the night sky, in a steady flux of about 10²⁹ ergs per day. This unceasing flow of free energy has kept the molecules on the planet in a mild but unceasing turmoil, during which the less stable combinations have incessantly been superseded by the more stable. Today, after 10¹⁰ years, what remains is mostly of extremely high stability, ranging from such minerals as granite to such dynamically stable forms as the mammal, a form older than the Alps and the survivor of several ice ages.
Looked at in this way, natural evolution and the emergence of forms such as the guinea pig, with its well-developed goal, are in no way unusual: they merely exemplify the fact that almost all state-determined systems tend to preferred regions. It is the exceptional system that does not show such preferred regions. Thus, the continuous state-determined system with equations2
shows nonconvergence only where div φ has the special value 0.2 Take φ's at random, and the system resulting is almost certain to show stability at preferred regions, within which it will behave in the goal-seeking way.
"Making a system that seeks a goal" is thus trivially easy: one forms a state-determined dynamic system at random (e.g., let it be specified by the spins of a coin). One is then almost certain to have a system that, like a guinea pig, will show that it is actively goal-seeking for some preferred state. It is true that the preferred state may well be meaningless or useless to the designer, but we should notice at this point that getting a machine to have some goal is no problem at all.
Achieving an Assigned Goal¶
Having disposed of this pseudoproblem, we can now consider the real, and difficult, problem. This arises when the designer not only wants the system to be goal-seeking but also wants it to seek some goal already specified. Air-traffic control systems are required to make collisions minimal, not maximal, and a physiological stabilizer of blood composition must have as its goal just those values that the human finds normal. Here the majority of stable states that might occur on random assembly are not acceptable.
When the system is small (designing a room thermostat, say), the designer needs no further general theory; he goes straight to the particular details. But when the system is of "cybernetic" size, he may still be uncertain of the next steps. I want to suggest in this paper that the general nature of the situation can be made much clearer if we apply what is already known in information theory.
The situation is most evident when a designer faces a heap of components from which he is going to construct his machine (but it is equally so when he faces a sheet of paper on which he will write a program). The point is that by his selection of the assembly or program he wants, from the set that includes both what he wants and what he does not, he transmits a message to the end product, and all the laws of communication are applicable. To design a thermostat that will hold 72°F is to transmit the value "72" to the machine. Consider the less trivial example of the designer who wants to allot values $-2, -1, 0, +1$, or $+2$ to the four coefficients $a, b, c, d$ in
so that the system will be stable. In this case the quantity of information that must be transmitted from designer to system is calculable. 5⁴ types are possible, of which 114 have the real parts of their latent roots both negative. In the worst case (if all values are equiprobable), the selection implies a transmission of log₂(625/114) bits, i.e., just under 2½ bits. Thus, in this example, the channel represented by Fig. 1 must be able to transmit at least 2½ bits (per act of design).

Fig. 1. From W. Ross Ashby, Setting Goals in Cybernetic Systems (1972).
The example is trivial: what matters is whether the principle is sound. If so, it will give us a deeper insight into problems that are anything but trivial.
Before we go further, however, we must notice a matter that might easily confuse us. Suppose some complex regulator accepts $m$ inputs Xᵢ (i = 1, …, m), data about aircraft at an airport perhaps, and emits orders, values on $n$ variables Yⱼ (j = 1, …, n), to the aircraft. The designer is then asked to design it so as to be a "good" traffic controller. The basic situation may be represented as in Fig. 2. How the outputs depend on the inputs is a relation F, the "transfer function" (in a general sense) of the system. The designer's task is to act so that the desired goal, selected from all the events (good and bad) that can occur at an airport, is transmitted to, and acts effectively as, a "good" F. This good F, it should be noted, comes from the set of all possible F's (not from the set of Y-values), so the transmission implied by the selection, along the channel C, is essentially independent of the transmission from X to Y. In this paper our attention will be focused on the quantities transmitted through such a channel as C.

Fig. 2. From W. Ross Ashby, Setting Goals in Cybernetic Systems (1972).
The situation perhaps can be made intuitively more vivid if the designer's task be formulated as that of conveying to the heap of components what he means by the phrase "a successful airport," in the operational mode of getting the heap, when assembled, to separate the "good" set from its complement. Similarly, the problem of designing a pattern-recognizer (for genuine dollar bills, say) may be regarded generally as one of trying to tell the machine, in its undeveloped state, what is meant by "genuine, dollar, and bill." Again, this flow of design-information is essentially distinct from the flow that occurs later when the finalized machine scans an actual piece of paper and emits a verdict. Similarly, in information retrieval (getting from a library, say, the documents relevant to "social feedbacks in universities"), the difficulty can be regarded as essentially one of getting across to the machine what is meant by "society, feedback, and university." In general, thinking of the design process as one of telling the machine what you want helps to make more evident the flow of information that is intrinsic to the act of designing.
Sponsor and Designer¶
If this thesis is granted, we are led inescapably to the deduction that exactly the same situation exists at the prior stage, represented in Fig. 3.

Fig. 3. From W. Ross Ashby, Setting Goals in Cybernetic Systems (1972).
Whence came the particular goal that the designer transmitted to F? Rarely is the designer himself the originator; more commonly, as in the example of the air-traffic control system, the goal comes from what I will call generically the sponsor. Again I argue, as before, that effects must have causes. Selection among the F's can be attributed to prior selection among the designer's possible goals. And this selection must be attributed to prior selection among the sponsor's possible goals. Further, it is axiomatic through information theory (though seldom stated explicitly) that complex effects, as "messages received," require at least as much complex causation as "messages sent." By as much we mean either the number of bits or, more simply, the number of possible messages. So with this axiom we can assert that there must be sufficient transmission, from sponsor to designer, for the selection made by the sponsor (among his possible goals) to reach the designer.
Often this requirement is so obvious that these remarks would appear to be unnecessary. But it also often happens that the requirement is not, in fact, adequately met. It is more than likely that at this meeting there are some designers who have been given nothing like sufficient information (channel B) about the sponsor's goal, yet who are expected to achieve high and appropriate selection in the system F. Sponsors must learn enough elementary cybernetics to appreciate when they are asking the impossible. The sponsor must learn, in fact, that he is as subject to Conant's First Theorem3 as every other would-be regulator. He is a regulator insofar as, with some goal in mind, he is trying to bring the designer, who might design all sorts of things, to accept his particular goal (for transmission to the system F). If the designer misunderstands and designs for G' instead of G, the sponsor must correct him and keep acting until the designer has the sponsor's G as his guide. This is an act of regulation and so is wholly subject to Conant's Theorem, which asserts that transmission of the corresponding quantity of information (through B) is absolutely unavoidable. Armed with this theorem, the designer can demand that the sponsor transmit sufficiently.
The only worker in this field who seems to have clearly understood and stated these requirements was the elder von Moltke, who, in 1858, founded and organized the German General Staff. The key principle he gave it was that of the "directive." In this method the senior (corresponding to the sponsor here) gives no orders to the junior. Orders were replaced by the rule: the senior shall take all necessary time to explain to the junior what the senior wants from the senior's point of view; then he shall leave the junior free to use all personal initiative and local knowledge to achieve the goal. Von Moltke evidently had the clearest intuitive understanding of these quantitative flows.
Sometimes, however, the sponsor need not transmit everything. What is necessary is that to the designer shall come sufficient information (as determining factors) to enable him to select his goal adequately. But the determination need not all come from the sponsor. A sponsor may well specify so much and then say "I delegate the rest to ——————." He may delegate it to a junior, who must then supply the rest. He may delegate it to the designer himself. The designer may himself delegate further, to the spins of a coin, perhaps. But in all cases the total of determination coming to the designer must be not less than the quantity required for the selection of his goal.
Such supplementation may take various forms. One well-known form occurs when the genes set the goal for the living organism. But, in the higher organisms, the goal is too complex to be transmitted through the genetic channel, so a part of the determination is delegated to the environment. Says the gene-structure to the kitten, "I have told you something about mice—now go out and get the finer details from the mice themselves." We call such supplementation "learning." And a learning machine is simply any machine that is specified only partly by its designer, who delegates the remainder of the specification to some "teaching" environment.
Quantities of Information¶
Once these general principles governing goals are clear, the rest is a matter of special techniques in special cases. But there is one aspect that I would like to mention, as I believe it to be of central importance. It concerns the case in which the goal is of very high complexity, as in artificial intelligence, high-order pattern recognition, and high-order regulation.
Without stopping to examine closely the idea of "complexity," I shall assume that a complex goal is one which has many parts and in which the required relations between the parts show high conditionality of part on part; when, that is, the whole goal is a nonreducible function of many variables. I want to stress that the difference in information content between the simple (reducible) and the complex (nonreducible) goal is enormous. The basic argument can be given clearly by using set theory. Thus: If each of n variables can take k distinct values, then the number of reducible (i.e., rectangular) relations between them is the number of rectangular subsets. In the space of kⁿ points, it is 2ᵏⁿ. But the number of relations in general, not restricted to the reducible, is 2^(kⁿ). Thus, specifying one of these relations (the subject3 of events acceptable as goal-achieving) requires, in the reducible case, $kn$ bits, and in the nonreducible case, kⁿ bits.
The difference may be triflng4 to the printer, but numerically it is out of this world. Suppose, for instance, that the traffic at an airport involves only 100 variables (probably an underestimate) and that each variable need be distinguished in only 5 degrees (again, a very moderate demand). If the goal is reducible, specifying it may demand up to 500 bits; if it is not reducible, the quantity rises to 10⁷⁰ bits, one bit for every atom in the universe! This fantastic leap is in no way exceptional; on the contrary, all that I have done in the last few years has shown that it is entirely typical. Allowing interaction commonly makes the informational content increase by vast orders of magnitude.
In these huge quantities we have a useful fixed point, and can keep some sense of proportion, by remembering Bremermann's limit.4 Because of the quantal coarseness of matter, nothing made of it, machine or brain, can process information faster than about 10⁴⁷ bits/p/sec.5 Take tons of computer and decades of time, and no feasible computation can handle more than about 10⁶⁰ to 10⁷⁰ bits. Thus, a generally complex goal, in the very moderate airport example just given, is already making demands quite beyond the achievable. I suggest that many of our troubles today, in our struggles with complex systems—especially in those researches that try to push into the really large and new—are basically ascribable to the fact that we are often attempting to handle quantities of information that are, by Bremermann's limit, actually unmanageable.
There are at least two ways in which the ideas and methods of information theory may help. First, even a rough approximation may suffice to warn us that we are attempting the impossible. Second, it may throw a quite unexpected light on the various strategic approaches to a problem. Here is an example that I encountered recently.
Suppose, as in Fig. 2, that the designer has to select the right function $F$ in this system with the $m$ inputs X₁, …, Xₘ. Now this system has an obvious transmission from the $X$'s to the $Y$'s, but transmission in information theory means more than the driving of electrons along a wire. Essentially, information theory is the science concerned with deviations from statistical independence. If the values of the $Y$'s are not independent of those of the $X$'s, then the ordinary "transmission through" occurs, but other deviations are possible. Thus the Xᵢ's may show deviations from independence (correlations, say) among themselves; then a transmission may be defined and measured between the Xᵢ's. In Ref. 5,5 T(X₁ : … : Xₘ) will not be zero. What effect will such correlation have on the quantity of information that must be handled by the designer?
Suppose T(X₁ : … : Xₘ) = 20 bits. Another way of writing this fact follows.
Now "2 to an exponent entropy" is effectively the number of independent values, equivalent to the correlated. So the expression says that the transmission of twenty bits cuts the effective number of states at the input to the fraction 2⁻²⁰. Next, take the fact that the number of mappings of $p$ input states into $q$ output states (i.e., the number of $F$'s from which the designer must select) is qᵖ. To select one $F$ may demand up to p log₂ q bits. If $q$ is fixed, the number of bits is proportional to $p$. But $p$ has been cut by 2⁻²⁰, i.e., to one-millionth. Thus, the transmission, among the Xᵢ's, of twenty bits does not just, for instance, subtract twenty bits from the designer's work: it cuts his work to a millionth part.
Having in mind this example and my other experiences of the last five years, I think it may reasonably be asserted that our most urgent need in artificial intelligence (and similar researches in highly complex systems) is that we be aware at every moment of whether the information content is dependent upon a multiplier or upon an exponent. Crude and elementary as this distinction is, without it a worker may struggle to achieve a ten percent saving in efficiency, unaware that, because of a wrong basic strategy, he is working at a level a million times too high.
Bibliography¶
- W. McDougall, Psychology (New York: Holt, 1912).
- W. R. Ashby, "The Set Theory of Mechanism and Homeostasis." In Automaton Theory and Learning Systems, ed. D. J. Stewart (London: Academic Press, 1967), pp. 23–51.
- R. C. Conant, "The Information Transfer Required in Regulatory Processes." IEEE Transactions SSC-5 (1969)6: 334–338.
- H. J. Bremermann, "Quantal Noise and Information." 5th Berkeley Symposium on Mathematical Statistics and Probability 4 (1967): 15–20.
- W. R. Ashby, "Two Tables of Identities Governing Information Flows within Large Systems." Communications of the American Society for Cybernetics7 1 (1969): 3–8.
Footnotes
- ↩
sic — "Borden" (for Burden) in the print (p. 33); preserved.
- ↩
sic — the print sets the left-hand side as xᵢ with no dot (p. 35); preserved. The 1981 reprint transcription gave ẋᵢ; the equation is the continuous state-determined system, so the fluxion is meant.
- ↩
sic — "subject" (for subset) in the print (p. 42); preserved.
- ↩
sic — "triflng" in the print (p. 42); preserved.
- ↩
sic — "bits/p/sec" in the print (p. 42), where g (gramme) is meant; preserved.
- ↩
the print has "(L969)" (p. 44); the footnote keeps the note's corrected year.
- ↩
the print has "Communications of American Society for Cybernetics" (p. 44); the footnote keeps the note's form.