Adaptiveness and Equilibrium

TL;DR. In 1940, a psychiatrist named W. Ross Ashby argued that "adaptive behavior" is not a special faculty of living things — it is simply what any dynamical system does as it moves toward a stable equilibrium, and organisms are just systems whose feedback keeps a handful of essential variables (temperature, blood sugar, structural integrity) inside survivable bounds. This reframes learning as re-equilibration rather than maximization, and it is arguably a better native objective for long-lived robots and self-maintaining agents than reward — an idea that is now computationally cheap to test and still barely explored.

The idea Ashby had in 1940

Ashby was working as a psychiatrist, and the paper opens from a clinical observation: what we certify people for losing is adaptiveness. Everyone agrees it is the central fact of animal behavior; nobody in 1940 could say what it is without invoking purpose, instinct, or some vital principle.

Ashby's move is to dissolve the mystery into dynamics. Take an organism plus its environment as one coupled dynamical system. Some of the system's variables are essential: if they leave certain bounds, the organism ceases to exist as an organism. Ashby's claim is an equivalence: a behavior is adaptive exactly when it is the behavior of a system returning toward a stable equilibrium of those essential variables. A cat pulled off balance rights itself; a mammal in the cold shivers and vasoconstricts; a person who touches a stove withdraws. In each case the observable "purposive" behavior is the trajectory of a state vector flowing back into a viable region. No homunculus needed — adaptation is a property of the vector field.

The state of the art this was written against matters. In 1940, behaviorism explained learning by stimulus–response chains, Cannon had just named "homeostasis" (1932) as a physiological curiosity, and there was no control theory of the brain — Wiener's Cybernetics was eight years away, McCulloch–Pitts neurons three. Ashby was, as far as I know, the first to say plainly that the mathematics of stability is the mathematics of adaptation, and to draw the corollary that systems which persist are, tautologically-but-usefully, those whose feedback structure keeps them viable. Persistence is selection at the level of dynamics.

Ultrastability: the double feedback loop

The 1940 paper is the seed; the full mechanism, ultrastability, came in Ashby's later work (a 1945 paper and Design for a Brain, 1952), but it is implicit here and worth stating because it is the actionable part.

An ordinary regulator has one loop: fast feedback between the reacting part and the environment. Ashby adds a second, slower loop. The essential variables are monitored against their bounds. As long as they stay inside, nothing changes. When one is driven outside — the environment has changed in a way the current feedback structure cannot handle — a set of step-functions fires: discrete, effectively random changes to the parameters of the reacting part. This rewires the fast loop. The random changes continue until, by chance, a configuration is found whose dynamics pull the essential variables back inside their bounds. Then the step-functions fall silent and the new configuration persists — because persisting is what stable configurations do.

Two features deserve emphasis because they are still unusual. First, the objective is a set, not a scalar: stay inside the viable region. There is nothing to maximize, and no incentive to overshoot. Second, structural change is gated by failure: plasticity fires only when the current dynamics are demonstrably inadequate. This is a principled answer to the stability–plasticity dilemma, given decades before it was named.

Why it could not work then

The 1940 paper is pure argument; there was no way to demonstrate it. To show ultrastability you need a multi-variable coupled dynamical system whose parameters can be automatically reconfigured and whose behavior you can observe over many adaptation episodes. In 1940 that meant custom electromechanics, and there was a war on.

It took Ashby until 1948 to build the Homeostat: four units assembled largely from ex-RAF surplus (magnet-driven coils, water troughs as integrators, uniselector stepping switches to implement the random step-functions). Roughly, each unit's uniselector had 25 positions, so four coupled units gave on the order of 25^4 \approx 3.9\times10^5 possible configurations — and the machine could try one configuration every second or two. The demo worked and became famous ("the closest thing to a synthetic brain," Time said, roughly), but the scaling wall was immediate and Ashby knew it: random search over configurations is exponential in the number of units. Ten units at 25 settings is \sim10^{14} configurations; a hundred units is \sim10^{140}. At one trial per second, the four-unit machine sat near the edge of what a human lifetime of watching could explore.

So the 1940 paper was unfalsifiable-in-practice: no simulation, no fast reconfigurable hardware, and no theory of directed structural search to replace the blind step-functions. The idea went into the cybernetics literature as philosophy plus a four-unit toy.

What changed

Three things, all mundane and all decisive.

Simulation is free. A single modern GPU runs physics simulators at millions of environment steps per second across thousands of parallel instances — very roughly ten orders of magnitude more "trials per second" than the Homeostat, on systems with millions of state variables instead of four.

The slow loop no longer has to be random. Gradient descent, evolution strategies, and meta-learning are all directed search over parameterized dynamics. Ashby's step-functions were random because 1948 offered nothing else; the architecture (structural change gated by essential-variable violation) survives even when the search operator is replaced.

The math of viability exists. Aubin's viability theory (1990s) formalizes exactly Ashby's object: the viability kernel, the set of states from which some control keeps the trajectory inside constraints forever. Hamilton–Jacobi reachability and control barrier functions make kernels and their surrogates computable, and safe-RL work already learns approximate viability sets for real robots.

A 2026 revival: viability as the primary objective

Here is the experiment the 1940 paper implies and nobody has quite run at scale.

Take a long-horizon embodied setting — a legged or wheeled robot deployed for weeks in simulation-plus-hardware, with battery, motor temperature, structural wear, and calibration drift modeled as state. Define essential variables e(x) \in \mathbb{R}^k and bounds \mathcal{B}, where x is the full state. Train two agents:

• Reward-maximizer (baseline): standard RL on task reward, with safety handled the usual way (penalties or constraints bolted on).
• Ashby agent: the primary objective is viability — maximize the probability that e(x_t) \in \mathcal{B} for all t, or equivalently stay inside a learned viability kernel — and tasks arrive as perturbations: disturbances that, if unaddressed, will eventually push essential variables out of bounds (an undone chore drains battery reserve margins, a blocked route raises energy cost of the patrol it must complete to stay charged). Task competence is emergent, not rewarded.

Keep from Ashby: the set-valued objective; the two-timescale architecture; failure-gated plasticity — the slow loop (fine-tuning, module recruitment, controller re-synthesis) fires only when the essential variables approach or breach their bounds, giving continual learning without catastrophic drift. Replace: random step-functions with meta-learned or gradient-based structural updates; analog units with a learned world model whose viability kernel is estimated by HJ-reachability or a learned classifier.

The measurable predictions: the Ashby agent should show better long-tail robustness (no reward-hacking trajectories that sacrifice integrity for score), graceful degradation under novel disturbances, and far less designer effort spent on reward shaping — at the cost of weaker peak task performance. Whether that trade is favorable for month-long deployments is precisely the open empirical question.

Descendants, vindication, and what is still open

The lineage is rich, which is itself evidence the 1940 framing was right. Friston's free-energy principle explicitly descends from Ashby: minimizing surprise about essential-variable states is a probabilistic restatement of staying in bounds. SMiRL (Berseth et al., 2019) showed that pure surprise-minimization in dynamic environments yields competent-looking behavior with no task reward — a direct, if partial, vindication. Keramati and Gutkin's homeostatic RL derives reward from deviation of internal variables from setpoints, unifying drive reduction with RL. On the control side, viability kernels, control barrier functions, and HJ-reachability are exactly Ashby's essential-variable bounds made computable, and they are standard in safety-critical robotics. Continual-learning work on gated plasticity and neuromodulation rediscovers the failure-triggered slow loop.

What has not been done, to my knowledge: a serious head-to-head between viability-as-primary-objective and reward maximization for general-purpose embodied agents over long deployments; a scalable learned analogue of ultrastability where structural search is directed but still gated by essential-variable violation; and any theory of what task repertoires emerge "for free" from viability pressure alone. Ashby's deepest claim — that adaptiveness needs no objective beyond persistence — remains untested at scale, and is now cheap to test.

Where to read it

The paper is "Adaptiveness and Equilibrium," Journal of Mental Science 86 (1940), doi:10.1192/bjp.86.362.478 (bibliographic details verified). It is short and readable. Read alongside: Ashby's Design for a Brain (1952, esp. the ultrastability chapters) for the full mechanism; Aubin's Viability Theory for the modern math; Friston's free-energy papers and Berseth et al.'s SMiRL for the two most direct modern descendants; and Keramati & Gutkin on homeostatic RL for the bridge to reward. Pickering's The Cybernetic Brain gives the historical texture, including the RAF-surplus Homeostat story.