Two parent designs are recombined into six offspring. You choose which ones look right. The zones on the drawing board decide which ones are allowed to exist.
Starting designs in three categories, plus anything you keep. Scroll sideways for more.
Drag a zone to move it and drag its corner handle to resize. With a zone focused, arrow keys move it and Shift plus arrow keys resize it. Click an offspring drawing to put it on the board.
A held group is copied from parent A unchanged, so evolution only explores the rest of the car.
Edit a gene directly before breeding. The board shows the result against the zones.
| Generation | |
|---|---|
| Candidates bred | |
| Discarded by a zone | |
| Closest misses shown | |
| Zones broken by parent A | |
| Zones broken by parent B | |
| Zones |
This page can be driven from code through window.sim (see the JSON block script#sim-manifest), or without code: load it with ?seed=1&generations=20&chooser=target&pick=2&autorun=1 and the result appears below.
Nobody can write a formula for "looks right." The algorithm does not try. It proposes variations and you choose. Each generation recombines the genes of your two picks and adds random variation, so your taste steers the search without ever being stated.
People have to fit. The roof has to pass under a limit. The driver has to see over the hood. Each zone is a pass or fail test. Candidates that fail are discarded before you see them, and the dots above the offspring show how many were culled.
Enlarge the occupant zone or lower the roof limit and watch the cull rate climb and the offspring converge. Hold a gene group from parent A to freeze a decision already made. If nothing valid turns up, the closest miss is shown and flagged.
Background: the A and B parent loop, the two offspring rows with different mutation strength, and constraint regions placed on the drawing follow the Shapeshifter user guide. Designer-in-the-loop genetic algorithms for car side views are studied in Cluzel, Yannou and Dihlmann, "Using evolutionary design to interactively sketch car silhouettes and stimulate designer's creativity," Engineering Applications of Artificial Intelligence 25 (2012), which encodes the silhouette as Fourier harmonics. This page uses a simpler genome of 20 proportions so each gene stays readable. The hood, cabin and trunk genes are measured from the axles, and roof length scales with glass height, so every gene combination decodes to a plausible car before your zones are applied. It illustrates the idea and does not reimplement either system.