Sailboat beats the wind
The buoy is straight into the wind, and a sailboat can't sail straight into the wind! Can 60 little boats figure out the trick? Press Play.
How the best half did, round by round
Checkpoints — the champion boat from each round, with its path. Tap to replay it on today's sea.
How it works
No sailing into the wind
Point straight at the wind and the sail just flaps. The boat stops. Point a bit to the side, and it zooms!
What a boat feels
Where the wind comes from, where the buoy is, how fast it goes, how close the shore is, and 5 feelers for islands.
A tiny brain picks a course
It says "go 45° right of the wind" or "45° left". An autopilot turns the rudder to hold it.
Discover zig-zag
Boats that switch sides before the shore get closer to the buoy. Their brains get copied. That's how tacking is discovered.
What just happened?
- In round 1 boats pointed at the buoy, stalled, or crashed into the shore.
- After a few rounds the best boats zig-zagged: left of the wind, then right, then left, all the way to the buoy. Real sailors call that tacking.
- Nobody taught them. Boats that happened to switch sides earlier did better, and their brains were copied.
- Some challenges barely bothered them, because they already think in "angle to the wind". Others (islands, a moving buoy, current) broke their plan, and they had to adapt.
For grown-ups: what's really going on
Neuroevolution of a small tanh network [14 → 8 → 1] over a population of 60, using the same genetic algorithm as the river boats (elitism, 30% crossover, spread Gaussian mutation).
- Sailing model: target speed follows a polar diagram: 0 inside the ±38° no-go zone, rising to full speed at 75–110° off the wind, easing off downwind. Speed eases toward the target (boats coast through a tack), plus leeway drift and optional current.
- Action space: the network outputs a desired course relative to the wind (−100°…+100°); a proportional autopilot steers to it. Choosing where to point rather than raw rudder made tacking evolvable: with direct rudder control the population never found it.
- Inputs: sin/cos of the buoy bearing and the wind direction relative to the heading, speed, distance, current, position across the channel and 5 ray-cast distances to islands and shores.
- Fitness: reaching the buoy scores 2 + time bonus; otherwise mostly the furthest progress along the start→buoy line, so one long diagonal earns something, but crashing (×0.5) or stalling in irons (×0.7) is penalised. The channel is deliberately narrow, which makes a single diagonal a poor strategy and rewards tacking.
- Robustness vs adaptation: because inputs are relative to the wind, wind shifts and gusts are a mild distribution shift; islands and a moving buoy change the task itself, so performance drops sharply and must be re-learned.