The Dog in the Loop: Control Systems, Perceptual Hierarchy, and Why Behaviour Makes Perfect Sense | Dog Smart Learning Centre
The Dog in the Loop: Control Systems, Perceptual Hierarchy, and Why Behaviour Makes Perfect Sense
By Oliver Ringrose - Dog Smart Training & Behaviour
By Dog Smart Training & Behaviour · · · 22 min read
By Oliver Ringrose - Dog Smart Training & Behaviour
I'm a systems engineer. For most of my working life, I've been obsessed with understanding how things work — and more specifically, why they work that way. I spent years building control systems: thousands of data points, sensors feeding back information, loops constantly reading error and applying correction. There's something deeply satisfying about watching all of that complexity settle into something stable and functional. Everything talking to everything else. The whole thing in balance.
But I learned something early on that changed how I thought about the work: you can have the most elegant control system ever built, running perfectly on paper, but if the environment surrounding it isn't designed properly, the whole thing falls apart. The system can't compensate for a fundamentally broken context. If the setup is wrong, no amount of tuning fixes it.
I still work with systems. I've just switched to organic ones. They're far more interesting. Far more delicate. They require exponentially more brainpower to understand — but the underlying principles? Strikingly similar.
How a Control Loop Actually Works
A control loop has one job: close the gap between where something is and where it needs to be. That target is the setpoint. The gap between setpoint and reality is the error. The loop reads the error and applies a correction.
Two components do most of that work.
The proportional element responds to the size of the error right now. Big gap — big correction. Small gap — small correction. But if the proportional response is tuned too aggressively, the system overshoots the setpoint, swings back, overshoots again, and oscillates either side without ever settling. Too weak, and the system crawls toward the setpoint but never quite arrives. Wrong gain and your loop is either swinging or stalling — neither of which looks like normal, stable behaviour.
The integral element handles accumulated error over time — the gap that hasn't yet been closed. It applies a steady ongoing correction even after the proportional element has done most of its work. Get the integral wrong and you either overcorrect and chase the setpoint in circles, or drag so far behind that the system is always playing catch-up.
There's also a derivative element — one that responds to the rate of change of error, essentially anticipating where things are heading. In many industrial processes it isn't needed, and I rarely used it. But in biological systems it's almost certainly present, built into reflexes and anticipatory behaviour, whether we label it that way or not.
The critical point about all of these: they only work if what they're controlling actually responds. A perfectly tuned loop controlling a seized valve will wind itself up indefinitely. The error accumulates. The correction escalates. The system keeps trying to fix something it has no power to fix. In engineering we call that integral windup. We design around it.
Dogs don't always have that option. And when they don't, we tend to call it frustration.
Wiener, the Nervous System, and the First Feedback Model of Behaviour
Before we go further, it's worth stopping to acknowledge where much of this thinking originates — because it didn't start in dog training, or even in psychology. It started in the Second World War, in the mathematics of anti-aircraft targeting.
Norbert Wiener was a mathematician at MIT working on the problem of predicting where a moving aircraft would be by the time a shell arrived at a calculated intercept point. The solution required the targeting system to model the aircraft's trajectory, correct continuously for error, and anticipate future movement. It required, in other words, a feedback loop with predictive capacity.
What Wiener noticed — and what he published in his 1948 book Cybernetics: Or Control and Communication in the Animal and the Machine — was that this was exactly what the nervous system does. Not metaphorically. Literally. The neuromuscular system is a feedback control system. It has sensors (proprioceptors, vestibular organs, spindle fibres), a comparator (comparing intended with actual movement), an error signal, and a correction output. The whole architecture is cybernetic.
Wiener's most striking evidence came from clinical neurology. Patients with damage to the cerebellum — a structure now understood to be central to motor control feedback — exhibited what he called purpose tremor, or intention tremor. When these patients reached for an object, the limb would oscillate: overshoot, correct, overshoot again, correct again, never quite arriving. Sound familiar? That's a control loop with too much gain and insufficient damping. It's a proportional controller that won't stop swinging. The biology was doing what the engineering does when it's tuned badly — not because the intention was absent, but because the feedback pathway was broken.
Wiener argued, and the subsequent decades of neuroscience have largely confirmed, that purposeful behaviour of any kind — not just reaching for a cup, but navigating a social situation, regulating arousal, managing a threat — requires this same architecture. A reference state, a sensed state, a comparison, an error, a correction. Over and over, continuously, at every level of the nervous system simultaneously.
This was a radical idea in 1948. Much of behavioural science at the time was dominated by stimulus-response models: something happens, something is emitted. Wiener's insight — and the insight of the cyberneticians who followed him — was that this model was incomplete. Behaviour isn't emitted. It's steered. The animal isn't pushed from behind by stimuli. It's pulled from ahead by reference states it's trying to reach. The feedback loop is the mechanism. The reference state is the goal.
Neural Currents: What the Signals Actually Are
William T. Powers — engineer, psychologist, and the architect of the theory we're about to spend some time with — took Wiener's framework and asked a question that most people had avoided: what, physically, are these signals?
Not metaphorically. Not abstractly. What is a "reference state" in terms of actual biology? What is an "error signal"? What is a "perceptual signal"?
Powers' answer, developed across decades of work and crystallised in his 1973 book Behaviour: The Control of Perception, was precise and unfashionable in equal measure: they are neural currents.
Magnitudes of neural activity. Actual rates of firing in actual neural pathways.
The perceptual signal is the neural current produced by sensory processing — the magnitude of activity representing what the organism currently experiences. The reference signal is a neural current produced higher in the hierarchy — representing what the system is trying to experience. The error signal is the algebraic difference between the two: reference minus perception. When there is no error, these currents are in balance and no corrective output is generated. When there is error, the output function — the muscles, glands, and neural effectors — acts on the environment to bring the perceptual signal back into alignment with the reference.
This is not metaphor. Powers was explicit: he was describing the actual physics of how the nervous system operates. The "goal" of a system isn't an abstract concept floating somewhere in consciousness — it is a magnitude of neural activity in a specific pathway, set by the level above it in the hierarchy, against which the current perceptual signal is continuously compared.
This matters for how we think about behaviour. When we say a dog is "trying" to do something, we are, in Powers' framework, quite literally correct. The system is running a difference calculation, in neural current, continuously.
The behaviour is whatever output closes that gap. The dog isn't choosing to behave badly. It is running a loop that is trying, in the only way available to it, to return its perceptual signals to their reference values.
Understanding this — really understanding it — changes what a behaviour problem looks like. It stops looking like wilfulness. It starts looking like engineering.
Hold my beer!
A simple example is someone carrying a glass of beer across a busy pub.
The person is not consciously thinking about wrist angle, liquid movement, grip pressure, shoulder position, body sway, visual tracking, and balance. Yet all of those things are being controlled. The perceptual condition being maintained is simple:
the beer stays in the glass.
When the person is standing still, the system is not inactive. Neural activity is already helping to maintain the relationship between the body, the glass, the liquid, and gravity. Opposing muscle groups are not simply switched off. Depending on the task, flexors and extensors around the wrist, elbow, shoulder, trunk, and legs may be co-activated or adjusted in relation to one another to maintain stability.
At this point, the perceptual signal — what is currently being sensed — closely matches the reference signal — the condition the system is maintaining. There is little error, so the corrective output is small.
Neural signals are present, but the output remains low and task-appropriate. The person is usually unaware of this level of control because the system is quietly maintaining balance.
When the person starts walking, the relationship changes.
The body moves, the glass moves, and the liquid begins to shift. Now the perceptual signal begins to deviate more from the reference condition. The system detects error. That error is transformed into output through the relevant control systems. The magnitude and timing of the output depend on the size and rate of change of the error, as well as on the gain and organisation of the control loop. Corrective action becomes stronger, faster, and more organised. The hand adjusts, the wrist compensates, the body slows, the feet alter direction, and the glass is kept within a safe range.
If someone knocks into them, the error becomes larger and faster. The beer surges towards the rim. The mismatch between reference and perception is suddenly magnified. The system responds with a larger and more urgent corrective output. The person may stiffen, lift the glass, twist the wrist, widen their stance, or step away. Much of this can happen before conscious thought has time to narrate it.
But this level of output would be inappropriate if the person were merely standing still. If the same strong correction were produced in a stable situation, the system would overcorrect. The glass would jerk. The beer would slosh everywhere. The person might throw beer over someone despite trying to prevent exactly that.
Drunkenness makes the example even clearer.
The perceptual reference may not change. The person may still be trying to keep the beer in the glass. But alcohol degrades the control system’s ability to maintain that relationship.
Sensory information may become less reliable. Timing may become slower. Error correction may become delayed. Muscular output may become less precise. The relationship between perception, error, and corrective output becomes noisier and less well regulated.
So the person may under correct, then overcorrect. The beer starts to tilt, but the correction comes too late. Then the hand jerks too far the other way. Instead of smooth, well-timed control, the system becomes clumsy and poorly scaled.
The reference condition is still there:
beer stays in the glass.
But the system is less able to bring the perceptual signal back into alignment with that reference. The problem is not that the person has stopped wanting to keep the beer in the glass. The problem is that the control loop has become degraded.
And the beer hopefully stays in the glass.
That is why drunken movement often looks exaggerated, delayed, unstable, or poorly scaled. The reference may remain, but the system’s ability to sense, compare, correct, and stabilise action has been compromised.
That is the point Powers is helping us see. The nervous system is not simply issuing commands to muscles.
It is continuously controlling perception by comparing the perceptual signal with the reference signal and acting on the error between them.
Output is not simply “on” or “off.” It is scaled through the organisation and gain of the control system, and it must be timed well enough to affect the perception being controlled.
Too little output, and the beer spills because the correction is too weak. Too much output, and the beer spills because the correction is too strong. Too much delay, and the correction arrives after the perceptual world has already changed.
And the beer does not stay in the glass!
I hope you called a taxi!
Control is not the absence of movement.
Control is the continuous adjustment of action around a perception the system is trying to keep within range.
Try this before you read any further.
Move your cursor — or your finger if you're on a phone — to keep the amber dot centred on the white ring. The ring moves on a complex, never-repeating path. Just track it for thirty seconds or so. Then hit Add disturbance and keep going.
Don't overthink it. Just do it.
What you just did was William T. Powers' tracking test — a deceptively simple demonstration he used to make one of the most important arguments in the history of behavioural science.
Watch the three graphs at the bottom. Your output graph — your hand movement — is constantly changing. It never settles. The disturbance graph, when active, oscillates wildly. But the error graph, the gap between the cursor and the target, stays relatively small and stable. You are continuously moving to keep it that way.
Here is the question Powers asked: what, exactly, is the behaviour?
Is it the hand movement? That varies constantly. It's different every time you run the test, different depending on how the target moves, different again when the disturbance is active. There is no fixed pattern to it, no consistent response to any consistent stimulus.
If behaviour is the hand movement, you cannot predict it, describe it, or explain it from the outside.
Or is the behaviour something else entirely — the maintenance of the cursor on the target? That stays stable. That is what you are producing, reliably, regardless of what your hand has to do to get there.
Powers argued it is the second one. The hand movement is not the behaviour. It is the output — whatever the system produces to protect the thing it is actually controlling. The behaviour is the control of a perception: the visual experience of the gap closing, of the cursor sitting inside the ring. Everything else is downstream of that.
When the disturbance is active, your hand and the disturbance become correlated — you compensate without knowing you are doing it, without being told, without a rule or a cue. The correlation score at the bottom will show this. But your error and the disturbance are not correlated.
The disturbance doesn't reach the controlled variable, because you won't let it.
You are not responding to stimuli. You are controlling perception.
Now ask yourself: when your dog adjusts its position, softens its posture, moves toward something or away from it — what is the behaviour? The muscle movement? Or the perception it is protecting?
PCT: A Hierarchy of What the System is Controlling
PI control explains the mechanics of how a loop corrects error. But it doesn't explain what the loop is trying to control in the first place, or why the same situation can produce completely different behaviour in two different animals.
For that, there is Perceptual Control Theory — PCT — developed by Powers across the second half of the twentieth century and most fully articulated in Behaviour: The Control of Perception.
PCT makes one central claim: behaviour is not a response to the environment. It is the control of perception. An animal isn't reacting to what's out there — it's acting to keep its experience of what's out there at a level it's trying to maintain. The behaviour is whatever it takes to do that, regardless of what the behaviour looks like from the outside.
And crucially, PCT organises this control as a hierarchy of levels, each one controlling increasingly abstract perceptual variables, each one setting the reference signal — the setpoint — for the level below it.
Here is what that hierarchy looks like in practice.
First Order: Intensity
At the base of the hierarchy are the simplest possible perceptual signals — raw sensory magnitudes. Light intensity. Sound level. Pressure. Temperature. The warmth of contact. The brightness in the visual field. These are controlled by the most basic feedback loops in the nervous system: reflexes, postural adjustments, the blink response.
A dog pressing into your leg on a cold morning is, at this level, controlling for temperature and tactile pressure — returning those perceptual signals toward their reference values. No cognition required. Just a basic loop, continuously correcting error.
Second Order: Sensation
The second level controls combinations of first-order signals — what we'd recognise as sensory qualities. Not just intensity, but the character of experience. A particular scent profile. A specific texture under the paw. The quality of a sound — not how loud, but what kind. This is the level at which a dog distinguishes between two different surfaces, or between the smell of one person and another, or between the sound of a familiar engine and an unfamiliar one.
This is also the level at which a great deal of olfactory work operates. When a puppy buries his nose in the ground, he isn't simply responding to smell — he is controlling for a specific quality of olfactory sensation, maintaining it, tracking it as it moves, working to keep that perceptual signal at its reference value. The nose-down, absorbed, oblivious-to-everything-else behaviour is the output of a second-order control loop doing its job.
Third Order: Configuration
The third level controls spatial arrangements — the configuration of sensory inputs relative to each other. Posture. Position. The layout of the body in space. The arrangement of things in the visual field.
This is where proprioception does most of its work. The dog that adjusts its weight, rounds its back, or drops its head on approach to something uncertain is running third-order configuration control — maintaining a body arrangement that keeps threat signals from escalating. The dog that holds a specific spatial relationship to its owner on a walk — not trained to heel, but choosing a particular distance and angle — is controlling for a perceptual configuration: where the human is relative to where the dog is, maintained continuously.
This is also the level at which a great deal of what we call "body language" actually operates. The display isn't emitted. It's held — an output that maintains a specific configuration against the perturbation of an approaching animal or person.
Fourth Order: Transition
The fourth level controls changes in configuration — movement, velocity, the way things shift over time. Not where things are, but how they're moving. Not a posture, but a gesture. Not a position, but a trajectory.
Recall Wiener's cerebellar patients, oscillating toward the cup they couldn't quite reach. That's fourth-order control gone wrong — the transition from here to there, the smooth arc of movement toward a goal, failing because the feedback pathway is broken. In an intact system, fourth-order loops control the quality of movement itself: the approach to another dog, the arc of a retrieve, the gait adjustment on uneven ground.
For a dog managing social encounters, fourth-order control governs the rate and manner of approach or withdrawal. The dog that moves in fast and hard, or that can't moderate its own velocity, isn't lacking a rule. Its fourth-order control system isn't finding its setpoint. The output keeps overcorrecting. The movement keeps oscillating.
Above the fourth level, the hierarchy continues — through the control of sequences (ordered events: the structure of a greeting ritual, the pattern of a hunt), relationships (the abstract connection between self and other, pack member and stranger), programmes (conditional, if-then chains of behaviour), and principles (the governing values that set the reference points for everything below).
Safety — in most animals, and certainly in dogs — sits high in this hierarchy. It has to. An unstable primary loop disrupts every loop operating beneath it.
When safety can't be maintained, nothing else can be properly controlled. The sequence loop can't complete its work. The relationship loop can't settle. The configuration loops oscillate. The whole cascade degrades from the top down.
This is why you cannot socialise your way past an unresolved safety problem. The secondary loop — the social information loop, the approach loop, the connection loop — cannot win while the primary loop is still swinging. You're trying to fine-tune a secondary controller when the primary system is still broken. It will not work. It cannot work. That's not opinion; it's systems logic.
When Hierarchies Collide
I see cascade conflicts regularly in social situations between dogs.
A dog that wants social information — scent, proximity, contact — is running a lower-order loop with a clear setpoint. But if the safety loop, operating higher in the hierarchy, is oscillating — unable to settle, swinging either side of its own reference — it overrides the social loop before it can complete its work.
The dog darts in. Gathers a fragment of scent. Safety loop fires. Pulls back. Error resets. Goes in again. Can't stay long enough. Leaves again. Returns. The social loop never reaches its setpoint because the safety loop above it won't permit stability.
And in doing so, the unstable system begins destabilising the other dog's safety loop too. Now there are two hierarchies in conflict — both winding up, both oscillating. What looked like a dog trying to say hello has become a fraught encounter that neither animal particularly wanted.
That's not a socialisation problem in the simple sense. That's a cascade conflict between hierarchical control systems. The upper loop has to stabilise before the lower loop can function.
Addressing the lower loop first — pushing the social interaction — is not only ineffective. It actively worsens the problem by adding perturbation to a system that is already failing to find its reference point.
Bandwidth, Automation, and the Cost of Learning
I was working recently with a young puppy with a genuinely impressive nose — already clearly a scent-led dog, absorbing olfactory information with total commitment. The problem wasn't that his nose wasn't working. It was that using it was consuming almost his entire cognitive budget.
In PCT terms, what was happening was this: his second-order olfactory control loops were new. They hadn't yet automated. Every unit of scent processing required active, conscious participation from higher levels of the hierarchy — because the lower loops hadn't yet learned to run themselves.
Think about learning to drive. When you're new to it, every element demands attention from the top of your own hierarchy downward. Mirror, clutch, gear, steering, road — all of it deliberate, active, effortful.
Holding a conversation at the same time is nearly impossible, not because you're incapable of conversation, but because your higher-order systems are fully occupied managing the lower ones.
As driving becomes fluent — as the lower loops automate and begin running reliably without top-down supervision — that processing load decreases.
The spare capacity returns.
That puppy's nose was doing exactly what a beginner driver does with the clutch. Every sniff required active, top-down processing. And while that was happening, the higher levels of his hierarchy had almost nothing left over for learned behaviour, for social responsiveness, for checking in. His owner went home with hidden treat games — not to tire him out, but to give those lower loops the repetition they needed to automate. To free the hierarchy above them to do something else.
The Phrases We Already Know
Most people are already more familiar with control systems than they realise.
"It's out of my control."
"I've reached my limits."
"I've got too much else to think about right now."
"I'm so wound up."
"You're winding me up."
We say these things about ourselves without a second thought. We understand instinctively what they mean — that systems have boundaries, that capacity runs out, that errors accumulate until something gives, that one person's instability can drive another's. We extend that understanding to ourselves, and usually to other people.
But somewhere along the way, we stop extending it to our dogs.
The dog that can't cope in a certain environment isn't being difficult — its loops are running against conditions they weren't designed to handle. The dog that escalates when blocked isn't being stubborn — it's winding up, exactly as any system does when the correction keeps failing and the error keeps accumulating.
The dog that shuts down when there's too much going on isn't being awkward — its hierarchy is fully committed, and the bandwidth simply isn't there. And the dog that winds up another dog? That's one unstable system doing what unstable systems do — destabilising everything around it.
These aren't character flaws. They're system responses. Cybernetic, hierarchical, feedback-driven system responses, operating precisely as Wiener described them in 1948 and as Powers formalised them in 1973.
We recognise them perfectly well in ourselves.
Maybe it's time we recognised them in our dogs too.
Dog Smart Training & Behaviour | Sevenoaks, KentForce-free. Systems-informed. Evidence-grounded.
Further Reading
Powers, W.T. (1973). Behaviour: The Control of Perception. Aldine.
Wiener, N. (1948). Cybernetics: Or Control and Communication in the Animal and the Machine. MIT Press.
