Mapping the Terrain

In the intro you borrowed the picture — a rolling terrain where height means fitness and every search can only walk uphill — and you watched a climber get marooned on a low peak. Borrowing a picture is fine for a first look. But to use the model, to run it on your own problems, you need to be able to draw the thing from scratch: to take “every possible version of some thing” and turn it into an honest terrain you can point at.

That’s this whole lesson. Good news up front — the landscape is built from only two ideas. Not five, not a dozen. Two. Get those two clear and the peaks, valleys, ridges and traps all fall out for free. So before we name anything, let’s make sure you can feel the first idea, because it’s the one that trips everyone.

That padlock is going to follow us through the lesson, because it’s the cleanest possible case of the first big idea: a vast set of possibilities, arranged so that one small change takes you to a neighbour. You just did the hardest conceptual move in the whole model — deciding who sits next to whom — without being told the rules. Let’s make it precise.

The horizontal axis: design space, laid out by similarity

The analogy. Think of a colossal hotel where every possible version of your thing gets its own room. Every possible padlock setting, every possible protein, every possible keyboard layout, every possible business strategy — each one has a room. Now here’s the rule the hotel is built on: rooms are assigned so that designs one small change apart are next-door neighbours. Change a single dial, swap a single letter, tweak one knob, and you’ve walked to the room next door. Change everything at once and you’re on the far side of the hotel. That hotel — every possibility, arranged by similarity — is the horizontal axis of a fitness landscape. We call the whole thing the design space: the complete set of all possible versions of the thing, spread out so that nearness on the map means similarity in the design.

The precise definition. The design space is the set of every possible configuration of whatever you’re studying, laid out so that two configurations sit close together exactly when one can be turned into the other by a small change (a single mutation, a single edit, a single adjustment). “Small change” is whatever counts as one step for your problem. The axis isn’t a number line you measure — it’s a map of possibilities, and the only thing that matters about a design’s position is who its neighbours are.

Worked example — the four-dial padlock. Back to the lock. Four dials, each 0–9, gives 10 × 10 × 10 × 10 = 10,000 possible settings — that’s your design space, ten thousand rooms. A “small change” is one click of one dial. So the setting 4-4-4-4 has neighbours like 4-4-4-5, 4-4-4-3, 4-4-5-4, 3-4-4-4 — every setting reachable by nudging a single dial by one. Notice two things. First, the space is finite but large: only four dials, yet ten thousand rooms, and real designs have vastly more knobs than four. Second, neighbours are defined by the moves you allow, not by the numbers looking similar — 4-4-4-4 and 4-4-4-5 are neighbours because one click connects them, while 4-4-4-4 and 7-1-9-2 are far apart even though both are just four-digit numbers.

Worked example — spelling out a word. Prefer letters to dials? Take every possible five-letter string as your design space (all 26⁵ ≈ 12 million of them). Let “one small change” mean “swap a single letter.” Then HOUSE is a neighbour of MOUSE (change the first letter) and of HORSE (change the third), but it is nowhere near PLANT, which needs every letter changed. Arrange all 12 million strings so that one-letter-apart words sit side by side, and you’ve built a design space you could, in principle, walk across one letter at a time.

When to use it

Reach for “design space” the moment you catch yourself thinking about all the ways something could be rather than the one way it currently is. Any time there’s a set of possible versions and a notion of “a small change from here,” you can lay it out as a design space: mutations of a gene, tweaks to a recipe, edits to a résumé, moves in a game, parameter settings for a model. The trade-off is that you have to commit to what counts as one small step — change that, and you rewire who’s a neighbour and reshape the whole terrain. Choosing the step size honestly is most of the skill; the rest of the picture just sits on top of it.

The vertical axis: fitness is height

Now the second — and final — idea. We have a flat hotel of possibilities. Time to give it altitude.

The analogy. Go back to the hotel and imagine raising each room up on a stilt whose height is how well that design works. A brilliant design gets a towering stilt; a hopeless one barely leaves the ground. Walk the corridors now and the floor rolls beneath you — up toward the excellent designs, down toward the awful ones. That rolling floor is the terrain. The height of the floor under any room is that design’s fitness.

The precise definition. Fitness is a single number attached to each point in design space that says how well that design performs at whatever it’s being judged on. In biology it’s reproductive success — how many surviving offspring a design tends to leave (survival and reproduction rolled into one). But the model doesn’t care what the performance measure is, only that each design gets one: a keyboard layout’s fitness could be typing speed, a protein’s could be how tightly it binds, a business strategy’s could be profit, a chess move’s could be its win rate. Whatever your yardstick, fitness is that yardstick’s reading, and on the landscape it becomes height. Tall = good, short = bad, and that’s the entire vertical axis.

Worked example — the padlock, scored. Suppose the lock’s actual combination is 7-2-7-2, and we score each of the 10,000 settings by how many dials it gets right. Then fitness runs from 0 (every dial wrong) to 4 (the jackpot, 7-2-7-2). The setting 7-2-7-2 is the single tallest tower on the landscape. 7-2-7-9 stands one notch lower (three dials right). 4-4-4-4 is down at height 0, in the flatlands. Now the terrain has shape: it rises as you click dials toward 7-2-7-2 and falls as you click them away. Every design has a horizontal address (which setting it is) and a height (how many dials it got right) — neighbours plus height, nothing more.

The misconception to kill: height is not complexity, and not size. It is deeply tempting to read the tall towers as the big, fancy, complicated designs and the short ones as the simple, primitive ones. Resist it hard. Height is only performance on your chosen measure. A dead-simple design can be the tallest peak on the landscape, and an elaborate, ornate one can sit in a valley because it works badly. A bacterium is vastly less complex than a human and yet is, by the measure of reproductive success in its niche, staggeringly fit — a towering peak. Height answers one question and one only: how well does this design do the job? Not how big, not how intricate, not how advanced. Just how well.

The vocabulary of terrain

Two axes give you a rolling terrain. Now we name its features — because once the landscape has hills and hollows, a small, precise vocabulary lets you say what you see, and these words recur every single lesson from here on.

Peaks. A peak is a design that is fitter than all of its immediate neighbours — a room whose floor is higher than every room next door. Stand on a peak and every single step you could take (every small change) leads downhill, to a worse design. That’s the definition of being stuck: nowhere to go but down. A peak that’s higher than its neighbours but not the highest thing on the whole map is a local optimum (you met the term in the intro) — locally the best, globally maybe mediocre. The four-dials-right setting 7-2-7-2 is a peak: every one-click change from it lowers your score.

Valleys. A valley is a low-fitness region — designs that work badly, sitting in the hollows between hills. Crucially, a valley isn’t just “bad,” it’s in the way: to get from one peak to a taller one, you usually have to descend into a valley first and climb back out the other side. Since an uphill-only search refuses to go down, valleys are the walls of the prison. In the padlock, a setting with zero dials right sits deep in a valley.

Ridges. A ridge is a connected run of high-fitness designs — a path along the high ground where you can move from neighbour to neighbour without ever dropping down into a valley. Ridges are precious: they’re the rare routes that let a search travel a long way across the landscape while staying up high, reaching distant designs it could never get to if the only high ground were isolated peaks. Think of it as a mountain trail that stays near the summit line instead of forcing you down and back up.

The global optimum. The global optimum is the single tallest peak on the entire landscape — the best design there is, full stop. Every other peak, however proud, is a local optimum by comparison. In the intro island, this is the summit marked with the dashed flag. The whole drama of the course is the gap between the peak a search actually reaches (some local optimum it climbed) and the global optimum (the flag across the valley it can’t cross). 7-2-7-2, the only setting with all four dials right, is the padlock’s global optimum.

Here’s the whole vocabulary in one place — terrain word on the left, what it means back in design terms on the right:

When to use it

Use this vocabulary as a diagnostic checklist whenever you’re staring at something that seems stuck or “optimised.” Ask: Is this a peak? (does every small change make it worse?) Is it the global optimum or just a local one? (is there clearly a better design somewhere the search can’t reach?) What valley is in the way? (what would you have to get temporarily worse at to escape?) Is there a ridge — a clever sequence of changes that stays high the whole way? The trade-off is that in a real, high-dimensional problem you usually can’t see the whole landscape, so you can’t be sure a taller peak exists — but knowing the vocabulary means you at least ask the right question instead of declaring victory on a molehill.

Wright, 1932: where the picture came from

The geneticist Sewall Wright introduced this picture in a 1932 paper with the gloriously unglamorous title “The Roles of Mutation, Inbreeding, Crossbreeding and Selection in Evolution.” He was wrestling with a hard problem — how populations could ever reach very good combinations of genes when getting there sometimes required passing through worse intermediate combinations — and he needed a way to think about it visually. So he drew a field of contour lines, like a topographic map, where each point was a combination of genes and the contours marked fitness. That sketch let biologists suddenly see something that equations had buried: that a population climbing toward better gene combinations could get trapped on a modest hill, blocked from a higher one by a valley of unfit intermediates. Nearly a century later the same picture has been borrowed by machine learning, protein science, engineering, and strategy — but it started as one geneticist’s attempt to draw a problem he couldn’t otherwise hold in his head. That’s the whole reason the model is worth learning: it makes an invisible difficulty visible.

Read a real landscape

You’ve built the terrain conceptually — now practice reading one. Below is a live fitness landscape. This time, don’t rush to drop a population; first just read the picture using everything from this lesson. Trace the horizontal axis: that’s design space, every possible design laid out so neighbours sit side by side. Trace the vertical axis: that’s fitness — the height of the terrain is how good each design is. Then find the terrain features by name: the peaks (points higher than their neighbours), the valleys between them, any ridge where high ground connects, and the dashed flag marking the global optimum, the single tallest summit.

Say it out loud as you look: “left-to-right is which design; up-and-down is how good; that bump is a peak, the best in its little neighbourhood; the dip beside it is a valley; the flagged one is the global optimum.” When those words come automatically, you’ve stopped borrowing the picture and started owning it.

The dimensionality caveat: the cartoon vs. the real thing

One honest confession before we go, because the model is often oversold. The landscape you’ve been looking at — a single wavy line, or at most a 3-D surface of rolling hills — is a cartoon. Real design spaces have as many dimensions as the thing has independent knobs to turn.

Why. Our padlock had four dials, so its true design space is four-dimensional — you’d need four axes just to place a setting, plus a fifth for its height. A protein 100 amino acids long has a 20-choices-at-each-of-100-positions design space; a genome has thousands of genes; a business has countless strategic levers. You cannot draw a 100-dimensional terrain. So we flatten it down to one or two dimensions for the page — which is exactly like drawing a mountain range as a single line on a chart. The line is a useful lie: it throws away almost all the directions you could actually move.

What survives the flattening, and what doesn’t. Here’s the reassuring part. The logic of the model — the part you’ll actually use — survives the cartoon completely intact, because it depends only on the two ideas that don’t care about dimension count: neighbours (a design still has neighbours one small change away, however many dimensions there are) and height (each design still has a fitness). And so the punchline survives too: an uphill-only search still climbs to a local peak and stops, whether the landscape has 1 dimension or 10,000. What you must not do is over-read the cartoon’s shape — count its peaks, eyeball the “distance” between hills, or assume valleys are as narrow as they look on a line. In high dimensions there are usually vastly more directions to move (which can mean more escape routes, since a wall in the 1-D picture may have a door you can’t draw) and the geometry gets genuinely weird. So trust the logic of the landscape; distrust the literal shape of any low-dimensional drawing of it.

Recap

You walked in able to borrow the fitness-landscape picture. You’re walking out able to build one from scratch and name every feature on it:

1. Design space is the horizontal axis — every possible version of the thing, laid out so that designs one small change apart are neighbours. It is not time and not progress; horizontal position is just an address, decided by who sits next to whom.
2. Fitness is the vertical axis — how well each design performs (survival and reproduction in biology, or any performance measure you choose), drawn as height. It is not complexity and not size: a dead-simple design can be the tallest peak, an ornate one a valley.
3. Two ideas run the whole machine — neighbours (the horizontal arrangement) and height (the vertical fitness). Every peak, valley, ridge and trap is just a consequence of those two.
4. The vocabulary of terrain: a peak is fitter than all its neighbours (every small change makes it worse); a local optimum is a peak that’s only best around here; the global optimum is the single tallest summit; valleys are the low ground a search must cross to reach a better peak; a ridge is a connected high path that never dips down. Sewall Wright drew it in 1932 to make an invisible difficulty visible.
5. The cartoon caveat: real landscapes have as many dimensions as the thing has knobs, so any 1-D or 2-D drawing is a flattened cartoon. Keep its logic (neighbours, height, uphill-only, local vs. global peaks — all dimension-proof); distrust its literal shape (peak counts and valley widths are drawing artifacts).

Where this goes next

You can now build a fitness landscape from nothing — design space along the bottom, fitness as height, and the full vocabulary of peaks, valleys, ridges and the global summit to describe it. You know the two ideas that run the whole machine (neighbours and height), the classic misreadings to avoid (the axis isn’t time, the height isn’t complexity), and the honest limits of the cartoon.

Next is lesson 3, Hill-Climbing — the engine that moves across this terrain. We’ve built the map; now we watch how a search actually travels it. You’ll see exactly why selection is an uphill-only, local, foresight-free walk, what “local” really means when you can only feel around your own feet, and why that single rule — never step downhill — is at once the source of all the model’s power and the reason it gets so reliably, so tragically stuck.