Natural selection as an online algorithm in rail operations

The person, the decade, the problem

Ronald Fisher published the fundamental theorem of natural selection in 1930, and in doing so gave a per-generation update rule for something Darwin and Wallace had described mechanically seventy-two years earlier without the mathematics to formalise it. Fisher tied the rate of increase in fitness to the additive genetic variance present in a population at that moment — not to some future optimum, not to a target fixed in advance, but to whatever variation currently existed when the environment took its measurement. George Price's 1970 equation made the bookkeeping exact: it decomposes the change in any trait from one generation to the next into the covariance between that trait and fitness, plus a transmission term. Neither man was solving an engineering problem. Both were solving the same structural puzzle: how can adaptation accumulate without foresight, when the thing doing the adapting never gets to see the data in advance and never gets to stop and retrain?

The answer, later formalised by Cosma Shalizi, David Harper and others from the 1990s onward, is that selection is an online algorithm in the strict computational sense. It takes one example at a time, in the order the environment delivers it, updates a running estimate — allele frequency — and discards nothing except by way of that update. There is no training set separable from a deployment set, because biology never gets a deployment phase in which the world agrees to hold still.

Track circuits do not hold still either

A network controller's board is a live stream, not a snapshot. Track circuit occupancy updates on the order of seconds. Rolling-stock telemetry — traction current draw, bogie vibration, brake pipe pressure — arrives continuously from units running in service, not from units parked for inspection. Weather feeds and maintenance-window schedules layer on top, each on its own clock, each capable of invalidating an assumption the others were relying on. None of this is a corpus. It is evidence arriving in arrival order, and the controller's running belief about the state of the railway is the online estimate Fisher's theorem describes for a population: reweighted by whatever has just been observed, never reset, never granted a pause to recompute from scratch.

The characteristic failure of this environment is well known to anyone who has worked a control desk during a rail defect event. A speed restriction gets applied after the defect has already propagated — after a rough riding report has become a broken rail, after a hot axle box alarm three stations back has become a wheel-flat detector trip closer in — rather than at the moment the first weak signal appeared. The restriction is correct. It is simply late, and lateness in a non-stationary system is not a rounding error. It is the interval during which a wrong belief kept being acted on.

This is exactly the failure mode selection cannot afford and engineered batch systems can fall into by design. A model trained once on historical defect data, however large that data set, is a corpus with a cutoff. It encodes the distribution of failures up to the date it was assembled and then meets a railway that keeps ageing, keeps being rained on, keeps having its rolling stock re-geared and its ballast re-tamped. The gap between when the defect signature entered the world and when the frozen model would have recognised it is precisely the gap Fisher's variance term measures for a population meeting a new predator: the lag is not a bug in the implementation, it is the structural cost of having stopped updating.

Where the Large World Model falls short too

Sensing helps, and it is worth being honest about how much. A control system that ingests live track circuit state and live telemetry is already doing better than a model trained once — it has a present. But a present scene is not a history. A Large World Model, in the sense used across this comparison, restores contact with arriving evidence within the horizon of a bounded scene: the current occupancy pattern, the current speed profile, the current weather cell over one section. That is an organism's perceptual moment, not its lineage's memory. It tells the controller what is true now. It does not, by itself, carry the provenance of why a particular axle box has been flagged three times this month, or how a wet-rail adhesion model was revised after the last autumn leaf-fall season, or which maintenance deferrals are already stacked against a section that is about to get colder.

The terminus is the third row, and rail operations is a plain demonstration of why. What a controller actually needs is not a better snapshot but a running, provenance-carrying belief: this section is under a caution not because a rule fired once but because three independent streams — vibration trend, rainfall accumulation, a deferred tamping job — have been separately reweighting the same belief for eleven days, and the belief can be traced back to each of them.

Two objections rail operations sharpens rather than settles

Natural selection is a spectacularly inefficient online learner. It has no gradient, no credit assignment, and it pays for every update in deaths. A control system that can afford to batch its inference should do so, because batch methods extract more information per unit of evidence than continuous reweighting ever will.

Conceded, and conceded fully. Selection is slow; fixing a single beneficial allele can take thousands of generations even under strong pressure. A well-designed batch process — overnight recalculation of a whole network's speed restriction table against the day's accumulated telemetry — will out-perform naive continuous updating on almost any efficiency measure. But efficiency and intake structure are different axes. The claim about rail operations is not that continuous update is a faster optimiser. It is that a system whose distribution of beliefs stops updating goes on being wrong for as long as it stays frozen, regardless of how efficient the freezing was. A once-a-shift batch recalculation is still, in structural terms, closer to the Large Universe Model than a model trained once at commissioning and never revisited — the question is how short the batching interval is, not whether batching happens at all. A fast online learner beats a slow one and is still online. The efficiency objection is real and should temper any claim that continuous re-fitting is free; it does not touch the claim that stopping is costly.

Rail infrastructure has plenty of things trained once and left alone — a permanent speed restriction imposed after a structural survey, a standard clearance gauge, a signalling headway calculated at design stage and not revisited for decades. If large parts of the system run on fixed rules, "everything continuously" is not actually what rail operations does, and the analogy overstates its case.

This is correct, and it is the same point Hox gene conservation makes for biology. A permanent speed restriction that has stood for twenty years is not evidence of update having ceased; it is evidence of a belief that keeps being re-tested by every train that passes over it, by every subsequent structural inspection, and that keeps returning the same answer. The moment an inspection finds movement in the underlying embankment, the restriction changes immediately — which is the tell that the apparent stability was earned by continual re-testing, not granted by neglect. The distinction matters operationally: a controller should be able to tell the difference between a rule that is stable because nobody has looked in ten years and a rule that is stable because it has survived ten years of looking. Only the second kind belongs in a system built on the intake structure this argument describes.

What the recurrence establishes

None of this shows that natural selection is a model for engineering a control system. It shows something narrower and harder to dismiss: whenever an operator has to act inside an environment that keeps moving — a population meeting a new pathogen, a controller meeting a defect that is still propagating — one-shot fitting produces the same failure, late correction of a belief the world had already overtaken. The lineage from a frozen corpus, to a bounded present scene, to every stream held open with provenance and decay, is not an argument invented for rail operations. Rail operations is simply where the argument shows up again, on a control desk, in the interval between when a rail starts to crack and when the restriction finally goes up.