Artificial Creativity: Novelty Search Evolutionary Art

Population

Population size 16

Mutation rate 0.12

Exploration ↔ Exploitation 0.50

Low = select for aesthetic quality only. High = select for novelty only, chasing shapes unlike anything seen so far.

Novelty Search State

Generation

Archive size

Avg. novelty (last gen)

–

How it works

Each artwork is an 8-gene genome g₀…g₇ ∈ [-1,1] decoded into a closed 3D parametric curve — this is the "creative act": a fixed decoding rule turning raw numbers into visual form.

x(t) = sin(fx·t + g1·π) · (1 + 0.35·cos(5·g2·t))
y(t) = cos(fy·t + g4·π) · (1 + 0.35·sin(5·g5·t))
z(t) = sin(fz·t + g7·π) · 0.8      t ∈ [0, 2π)

Each genome also gets a behavior descriptor b (spread on x/y/z, mean hue, path complexity). Novelty is the average distance, in that behavior space, to its k=5 nearest neighbors among the current population and an archive of past standout shapes:

novelty(i) = (1/k) · Σ dist(b_i, b_j)   over kNN(i)
combined  = w·novelty_norm + (1−w)·quality_norm

This is the core idea behind novelty search (Lehman & Stanley, 2011): reward being different from what has already been tried, not just being "good" by a fixed metric — because chasing a fixed quality score alone tends to get evolutionary search stuck on one safe local optimum, exactly the trap creative systems need to escape.

Quality here is a simple aesthetic heuristic — radial symmetry of the curve plus color variety — not a claim about universal beauty.
Each generation: the single best-scoring genome survives unchanged (elitism), the rest are bred by tournament selection (size 3) + uniform crossover + Gaussian mutation at the rate you set.
The top ~30% most novel shapes each generation are added to the archive, so the search keeps expanding its notion of "already seen" instead of only comparing against its current population.

One cell = one genome's artwork

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.