Fitness Landscapes

In 1932 a geneticist named Sewall Wright drew a picture that changed how people think about improvement of every kind. Lay out every possible version of a thing — every genome, every wing shape, every product, every strategy — as points on a vast map, arranged so that similar versions sit close together. Now add a third dimension: height, where taller means fitter — better at surviving, reproducing, working, winning. What you get is a rolling fitness landscape of peaks, valleys and ridges. And here is the whole idea in one line: evolution, and any trial-and-error search, can only walk uphill.

That sounds harmless until you follow it to its brutal consequence. A process that only ever steps uphill is a local search with no map and no foresight. It climbs whatever hill it happens to start on — and then it stops, stranded on the summit of that hill, even when a far taller mountain is plainly visible across the valley. It can’t get there, because the only route runs downhill first, and going downhill means getting less fit, and selection won’t take that step. This is the model that explains one of the most counterintuitive facts about the living world and the built world alike: why “good enough” so reliably beats “best,” and then refuses to budge. The human eye has its photoreceptors wired in backwards, casting a literal blind spot, because the ancestral eye climbed a hill it could never climb back down. The nerve that controls your voice detours all the way down into your chest and back up — a design a first-year engineer would flunk — because evolution was trapped on a local peak. Neither flaw is a mistake. Both are summits.

This is an advanced course, and it deliberately generalizes far past biology. The reason fitness landscapes earn a place in your latticework is that the exact same map governs machine-learning training (gradient descent is hill-climbing wearing a lab coat), engineering and design, business strategy (a firm perched on a local peak, disrupted by a rival standing on a different mountain), skill acquisition, and the deepest trade-off in all of search: explore versus exploit — keep climbing the hill you’re on, or jump off to look for a higher one? You’ll learn what makes a landscape smooth (one hill, easy) or rugged (many peaks, full of traps), what it takes to cross a valley (mutation, drift in small populations, recombination), and — tying straight back to the Red Queen — what happens when the landscape itself won’t hold still: when your rivals are climbing too, the terrain deforms under your feet, a conquered peak subsides, and there is no final summit to reach.

This course assumes you’ve met natural selection — variation, selection, heredity, and what fitness really means — because the landscape is just a picture of what selection does once you let it loose on a terrain of possibilities. It pairs with the Red Queen effect, whose moving-target insight becomes, here, a landscape that reshapes itself. Across the lessons you’ll drive an interactive landscape: drop a population, watch it hill-climb and get stuck, crank the ruggedness from one gentle hill to a jagged range, fire a mutation to leap a valley to a higher summit — and switch on a shifting terrain that sinks the very peak you just conquered. By the end, “it climbed the nearest hill and stopped” will be something you see in evolution, in code, in companies, and in your own stalled progress — and you’ll know the handful of moves that get a search off a local peak.