Adaptive control in astronomy
The problem that gave adaptive control its name was an aircraft, not a telescope. In the 1950s, engineers designing autopilots for high-performance jets found that no single gain set worked across the flight envelope: the same control law that damped oscillations at low altitude and low Mach number would either overreact or barely respond at 60,000 feet, where the air is thin and control-surface authority collapses. Whitaker's MIT rule, published in 1958, gave a way to adjust controller parameters against a reference model in real time. Kalman sketched the self-tuning idea the same year. The X-15's MH-96 controller flew this logic from 1961, holding a constant limit-cycle amplitude from runway to 350,000 feet by continuously adjusting loop gain as aerodynamic effectiveness fell away almost to nothing. The plant would not stay still, so the law that governed it could not either.
Astronomy inherited the same structural problem without borrowing the name until much later, and it inherited it in its sharpest form: a plant that changes at a rate no fixed schedule can enumerate in advance, because the plant is the sky itself, and the sky does not file a flight plan.
The fading transient
A survey astronomer running a wide-field time-domain programme wakes up to a broker's overnight digest: several hundred candidate transients, cross-matched against known variable stars, moving objects, and galactic coordinates, each with a machine-generated classification and a confidence score. Somewhere in that list is a kilonova, a tidal disruption event, or a supernova caught days before peak, in each case a physical process that decays on a timescale of hours to weeks and that can only be characterised — its spectrum taken, its light curve densely sampled, its host galaxy typed — while it is still bright enough to observe. The window is not negotiable. A transient that fades below the follow-up instrument's sensitivity before a decision is made is not a small loss of signal-to-noise; it is a total loss of information, permanently.
The naïve failure looks like a scheduling problem: not enough telescope time, not enough people to inspect the candidates. It is really an intake problem. The astronomer's belief about which of the hundred candidates deserves the one available hour of spectroscopic time is built from streams that are still running as the decision is made — the survey's own photometric alerts arriving every few minutes, a transient broker re-scoring the object as more photometry accumulates, an independent group's follow-up report changing the classification, an archival plate from decades earlier ruling out a foreground contaminant. A decision made against yesterday's classification, however well-founded yesterday, is a fixed law applied to a plant — the transient's brightness, its class probability, its urgency — that has already moved.
Fixed schedules, scene estimates, running estimators
This is exactly the distinction control theory drew between a fixed regulator and an adaptive one, mapped onto three ways an intake pipeline can be built.
A fixed-parameter approach treats the classification pipeline as something trained once, on a labelled archive of past transients, and then run unchanged against every new alert. Its coefficients were fitted to a corpus of already-classified objects; the sky generating tonight's alerts keeps drifting away from that training distribution as new instruments come online, as survey cadence changes, as previously rare classes — fast blue optical transients, say — turn out to be less rare than the archive suggested. This is a controller identified once and left open-loop.
A scene-bound approach does better within a single night. It ingests the current batch of alerts, builds a working model of what is astrophysically plausible tonight, ranks candidates, and discards that working model once the batch is processed, rebuilding from scratch tomorrow. This is closer to gain scheduling indexed by a snapshot than to genuine estimation: the parameters are re-fitted per episode, but nothing persists across episodes except what was baked into the original training.
The position beyond both is an estimator that never terminates: one that carries a belief about each active transient's class, magnitude, and rate of decline, continuously updated as new photometry, new spectra, and new broker re-classifications arrive, each contribution tagged with when it arrived, from which instrument, at what confidence, and how much it should be discounted as it ages. This is Åström and Wittenmark's self-tuning regulator, generalised past a single plant to an open population of them, with the recursive identification running indefinitely rather than resetting at each session boundary.
The 1985 objection, and its answer
The temptation is to treat "update on everything, continuously" as strictly better, and astronomy's own broker ecosystem shows why that is false. Automated re-classification pipelines that weight every new data point equally are vulnerable to exactly the failure Rohrs and colleagues demonstrated for adaptive control in 1985: a scheme provably stable under its design assumptions can be driven unstable by small unmodelled effects — in their case, high-frequency dynamics and sensor noise the design had not accounted for.
A broker that reclassifies an object every time a new alert lands is not more informed than one that waits. It is an oscillator with a very short period, chasing artefacts — a cosmic-ray hit, a satellite streak, a bad flat-field — that a slower, more conservative pipeline would have filtered out by simple accumulation.
This is a real and recurring failure mode in survey pipelines: an unguarded update rule that treats every incoming point as equally informative will flip a classification back and forth on the strength of a single noisy detection, and a telescope allocated against that flip-flopping estimate is worse served than one allocated against a stable, if slightly stale, fixed catalogue. The control-theoretic answer transfers directly rather than by analogy. Robust adaptive schemes survived the 1985 counterexamples by adding projection — bounding parameter estimates to a physically plausible convex set, so a transient cannot be re-estimated as brighter than its instrument's saturation limit implies — and by adding dead-zones and discounting, so an update below a confidence threshold, or from an instrument known to have a calibration fault that night, does not move the belief at all. A survey pipeline that discounts a re-classification because it came from a broker with a known systematic that week is doing the same bookkeeping. Guarded revision, not raw throughput, is the thing that works.
The excitation problem, and why provenance answers it
The second constraint is sharper still, and astronomers meet it every clear night. An estimate only improves with informative data, and a transient sitting near a survey's detection threshold, observed only in one filter, under poor seeing, is not informative about its true spectral class no matter how many such observations accumulate. This is the astronomical form of the persistency-of-excitation problem: volume of alerts is not the same as identifiability. A hundred more g-band-only detections do not tell you whether the object is a Type Ia supernova or a superluminous one; they tell you it is still there.
Feldbaum's dual-control tension — that probing a plant to learn it degrades the control you are delivering meanwhile — shows up as the literal cost of spectroscopic follow-up. Every hour spent confirming the class of a middling candidate is an hour not spent monitoring a stronger one, and every hour spent monitoring is an hour not spent probing an ambiguous one. There is no free way around this; it has to be budgeted, not wished away.
What answers it is not more data but provenance: an estimate of a transient's class that carries, alongside its point value, a record of which filters, which epochs, and which instruments contributed to it, and therefore a record of which directions in parameter space the recent data could and could not have constrained. A pipeline that knows it has not yet been given a spectrum, and flags the classification as photometric-only pending confirmation, is doing something a pipeline with a superficially higher-confidence score and no such bookkeeping cannot: it can tell the astronomer which candidates are worth spending the scarce spectroscopic hour on precisely because it knows what it does not yet know.
Where scheduling still wins
Not every regime in astronomy needs this machinery. Fixed cadence surveys observing a known variable-star field, or calibration schedules for a stable instrument, are closer to gain-scheduled flight control: the regimes — seasonal visibility, lunar phase, standard filter sets — are enumerable at design time, and a precomputed observing schedule can be validated in advance the way a certified autopilot's gain tables are. Adaptive intake earns its cost only where the regime space cannot be closed beforehand — where the object generating tonight's alert did not exist in the training archive, and the streams describing it are still running as the decision gets made. That is most of transient astronomy now, and it is why the discipline's brokers, however imperfectly, are already reaching for exactly this shape of solution.