hr-studioSearch · 05
A computational design primer · Evolutionary Search

A population learns the landscape.

Every strategy so far searched alone. Evolution’s move is to search as a crowd: score everyone, let the fit reproduce with variation, repeat. No individual knows anything — but selection quietly copies good discoveries through the population until the crowd itself becomes a map of where the summits are. Watch the cloud above find its mountain, generation by generation.

Scroll to begin
i.

The loop

One generation, slowed to its three beats: score every design against the objective, select — the unfit fade, the fit stay — and breed, each survivor spawning children that are copies plus a wobble of mutation. That’s the whole algorithm. Everything else is consequences.

Step through the beats and watch one generation turn.

Score → select → breed · that’s all of it
ii.

The cloud climbs

Run the loop and the population behaves like a creature: it drifts uphill, pools in basins, and pours itself onto the best peak it has found — because selection is a channel through which one lucky discovery becomes everyone’s starting point. One climber succeeded a quarter of the time here; the crowd almost always does.

Watch best and average fitness rise together.

Selection turns one discovery into everyone’s
iii.

The mutation dial

This landscape hides its true summit — a narrow spike on the flank of a broad, seductive hill. Turn mutation down and the population commits early, freezing on the big hill before anyone stumbles onto the spike. Turn it up to a healthy wobble and the cloud keeps enough restlessness to find it. Turn it to a boil and the cloud never settles at all — its best find is a lottery ticket, not a design language.

Try all three temperaments on the same trap.

Too little commits · too much boils · enough discovers
iv.

Crossover assembles

Mutation only wobbles; crossover recombines — a child takes some genes from each parent. On problems whose answers have separable parts, that’s a superpower: one parent has solved the ridge in x, another the ridge in y, and one of their children inherits both halves and lands where neither parent could reach. Evolution doesn’t just search; it assembles partial solutions.

Breed the two specialists and watch a child hit the crossing.

Two half-answers, one whole child
v.

Evolve the tower

The payoff: evolution as a designer. A population of twenty-four towers — the four-parameter vector from the start of this part is now a genome — bred against an objective you choose. Watch the family portrait converge generation by generation: score for floor area and slabs take over the gene pool; efficiency breeds mid-rises; slenderness breeds spires. The champion stands on the right. Nobody drew it.

A gene pool of buildings · the champion emerges
Score, select, breed: three beats, and a crowd of guesses becomes a lineage of designs. Evolution is search that remembers — but it optimizes whatever you score, one number at a time. Real briefs keep several scores at once, and that is the final chapter.