Ashby, Design for a Brain — ultrastability

Type: kb/sources/types/source-review.md · Tags: foundations, computational-model

Source: W. Ross Ashby, Design for a Brain: The Origin of Adaptive Behaviour, 2nd ed. (Chapman & Hall, 1960), Chapters 7–8. Read from the full-text PDF at https://ashby.info/Ashby%20-%20Design%20for%20a%20Brain%20-%20The%20Origin%20of%20Adaptive%20Behavior.pdf. Ashby numbers sections as chapter/section (e.g. 7/7); citations below give the section and the printed page.

Capture note: this is a source review, not a snapshot. The book is a copyrighted monograph and was not captured verbatim; the extraction below is authored from Chapters 7 ("The Ultrastable System," printed pp. 80–99) and 8 ("The Homeostat," printed pp. 100–121), which are the chapters that carry the ultrastability argument. Quotations are short and cited.

Key Points

Two loops, not one

An ultrastable system has "a motor output to the environment and two feedback loops" (7/5, printed p. 83). Ashby distinguishes them by role:

The first loop … consists of the ordinary sensory input from eye, ear, joints, etc., giving the organism non-affective information about the world around it. The second feedback goes through the essential variables … it carries information about whether the essential variables are or are not driven outside the normal limits, and it acts on the parameters S. The first loop plays its part within each reaction; the second determines which reaction shall occur. (7/5–7/6, printed pp. 83–84)

The first loop is ordinary regulation, operating within a way of behaving. The second is slower and operates on the way of behaving, by changing the parameters S that fix which behavior the reacting part R produces.

Essential variables and viability

The essential variables are those that must stay within physiological limits; Ashby draws them as "a dial with a pointer, and with two limit-marks, to emphasise that what matters about the essential variables is whether or not the value is within physiological limits" (7/3, printed p. 81). Viability is a boundary condition, not a graded score — the system registers in-limits or out-of-limits, and nothing finer.

How changes are tried

S is realized by step-functions: variables that hold constant and then jump (7/13, printed p. 87). A step-mechanism has critical states whose occurrence makes it jump to a new value (7/18, printed p. 91).

The Homeostat makes the generator concrete, and it is the decisive detail. Its uniselector values "were deliberately randomised by taking the actual numerical values from Fisher and Yates' Table of Random Numbers" (8/2, printed p. 103). When a unit's essential variable leaves its limits, the step-functions change to three new values, and Ashby is explicit about what those values are:

These new values have no special relation either to the previous values or to the problem in hand—they are just the values that next follow in Fisher and Yates' table. (8/3, printed p. 104)

Trial-and-error is not a second-rate method here but a necessary one, because the environment is a Black Box whose input–output relations can only be elicited by acting and observing the result (7/4–7/5, printed pp. 82–83).

What causes a configuration to be retained

This is stated as a rule, and it is the sharpest thing in the chapter:

If the trial is unsuccessful, change the way of behaving; when and only when it is successful, retain the way of behaving. (7/7, printed p. 84)

Operationally: when the essential variables are outside their limits, no state of S is equilibrial, so S must keep changing; when they are all within limits, every state of S is equilibrial, so S stays put (7/7, printed p. 84). Retention is therefore negative and structural — a configuration persists because nothing is left to displace it, not because anything scored it well. Ashby's veto-theorem argument concludes that the whole can only rest where both the essential variables and S are at equilibrium, so "if it goes to an equilibrium, the equilibrium will always be found to be an adapted one" (7/11, printed p. 86).

He further argues the second feedback is necessary: any system with essential variables and limits that adapts by testing behaviors "must have a second feedback formally identical (isomorphic) with that described here," for brains "living and mechanical" alike (7/8, printed p. 85).

What Ashby does not supply

Stated plainly, because the gaps matter more than the correspondences:

  • No explicit candidate search. There is no generator that proposes changes targeted at the problem. The Homeostat's new parameter values are drawn from a random-number table and bear no relation to the difficulty at hand (8/3, printed p. 104). Nothing selects which aspect to modify or what kind of change to attempt.
  • No evaluator or oracle. Nothing scores, ranks, or compares candidates. The only signal is the binary in-limits/out-of-limits state of the essential variables, and it does not evaluate a candidate — it merely fails to stop the reorganization. There is no representation of why a configuration is good.
  • No self-representation. The ultrastable system contains no model of itself. S is a set of parameters that get jogged, not structures the system inspects, reasons about, or acts through as a representation of its own organization. Ashby's mechanism is causal and material throughout, and this is precisely what distinguishes it from computational reflection.

Relevance to the KB

Ashby is the cybernetic precedent for the second loop — an adaptive system that changes its own organization when ordinary regulation fails, through a loop distinct from the one that regulates. The two-loop structure is exactly the distinction between operating a system and modifying it that the self-improving system definition builds on. Under that definition the Homeostat classifies as non-reflective, direct viability-driven self-improvement. It is not an instance of the proposal-selection improvement loop: the machine's one evidence-responsive transition — out of limits, jump — both discards the incumbent configuration and produces its successor, so rejection is not an operation distinct from generation, and persistence is equilibrium rather than a verdict. The loop's functions collapse into a single trigger, which is that note's exclusion criterion.

The Homeostat does admit a functional variation–selection–retention reading — configurations vary, viability determines whether variation continues, the survivor persists — but that mapping onto the loop vocabulary is an analyst's reconstruction, and it marks the floor rather than the archetype: variation is a random draw, the viability boundary is one-bit and ranks nothing, and retention is equilibrium — the absence of a force to change further. Ashby did not decompose adaptation into generator and oracle, and translating him as though he had would credit him with distinctions he explicitly did without.

Two things follow, and the second is the load-bearing one. Ultrastability shows how little machinery adaptation requires, and thereby what the extra machinery of a search-and-evaluate architecture is actually buying. And it shows what a self-representation buys, which is the line between this and a reflective self-improving system. On the broad definition the Homeostat is itself a self-improving system — its own organization changes under direct viability-driven adaptation, with no human anywhere — occupying the category's minimal, non-reflective, fully autonomous corner. The Homeostat genuinely retains — the surviving configuration persists and steers later behavior, which is what makes the retention operative — but what it retains is an opaque parameter setting rather than a representation. Nothing in the mechanism can read that setting, say why it was good, criticize it, or carry it to the next problem, so the process runs indefinitely and nothing compounds. Ashby is the clean demonstration that adaptation can retain without accumulating, and that routing an improvement pathway through a self-representation is what makes retention addressable.

The gap is most instructive at retention. Ashby's retention criterion is negative — a configuration survives because it stops the reorganization, not because an evaluator endorsed it. That is a real alternative to oracle-gated retention, and it is worth holding next to the oracle-strength spectrum, which grades evaluators but does not contemplate an architecture with no evaluator at all. It also contrasts instructively with the Gödel machine, the opposite extreme: retention licensed only by proof. Ashby retains anything that stops hurting; Schmidhuber retains only what is proved to help. The contrast is architectural, not a difference of gate strength: the Gödel machine runs an explicit proof-mediated proposal-selection loop, while the Homeostat has no gate at all.

Finally, Ashby is a clean negative case for reflective system. The ultrastable system adapts, modifies its own organization, and is causally coupled to its own viability — and it is still not reflective, because it has no self-representation. It is the sharpest available demonstration that adaptation does not imply reflection, and that the definition's causal-connection-through-a-representation requirement is doing real work rather than restating "the system affects itself."

Open Questions

  • Whether Ashby's negative retention criterion (persist because nothing displaces you) has a useful analogue in artifact-based systems, where retention is usually positive and authored.
  • Whether the random generator is essential to ultrastability or merely the simplest sufficient one — Ashby argues the second feedback is necessary (7/8), but makes no such claim about how trials are generated.

Relevant Notes: