KC Neural Racing Speedway

LiveLoading…

Progress by track

Milestones

    0:00.0

    Space plays/pauses · ←/→ previous/next race · 1–9, 0 pick a car (Shift for 11–20) · click a car to inspect it · drag to pan, wheel to zoom, right-drag (or Q/E) to turn and tilt, double-click for the whole track

    –—

      The 20 teams

      Every car has the same sensors and controls; the teams differ in how they learn (five kinds of evolution, gradient-based reinforcement learning, and combinations of them) and in their brain sizes and settings. Ranked by Grand Prix points over each team's last 20 Grand Prix races. Only finishers score: the winner gets a point for every car in the race, down to 1 for last place, and a car that doesn't finish gets 0. The bars show those races, oldest first; gray ticks are races it didn't finish. Hover a bar for that race. Click a team to see its fitness score over every generation. The greener a training stat's box, the better it is compared with the other teams.

      How it works

      Senses · 65 inputs

      12 distance rays (wall or car, closing speed), the most open direction ahead (the longest clear distance within 60° either side, checked every degree near straight ahead, and its angle), speed, sideways slide, spin, rpm, engine on, gear, pedal and steering positions, wheelspin, contact flags, the direction of the road 20 m and 50 m ahead, and its 4 memory values from the step before. No map coordinates, no track ID.

      Controls · 9 outputs

      Clutch pedal, gas pedal, brake pedal, steering wheel and a shift lever (up · hold · down), plus 4 memory values it hands itself for the next step (so it can count, e.g. hold the clutch in for a second). Braking hard uses up the tires' grip, so brake before the corner, not in it.

      The car

      Rear-wheel drive, 5-speed manual plus reverse. The clutch slips, so pulling away needs revs and a gradual release. Shifting needs the clutch pressed or the gearbox grinds. Below 450 rpm the engine stalls; holding the clutch in for 1 second restarts it.

      Crashes

      Hitting a wall or another car faster than about 16 mph wrecks the car, and it stays where it stopped as an obstacle.

      Training

      Each generation races on 2 of the 9 training races by default: The Bowl (a 1.9 km circle) and the Speedway Oval (2.5 km), each as a 1-lap and a 3-lap race in both directions, and the Half-Mile Drag Strip, which ends at the half-mile line (Settings → Tracks per generation). Every team's drivers race in groups, one driver per team, and each driver's score is its average over the tracks. Then each team learns its own way: the genetic algorithm, NEAT and MAP-Elites breed new brains from the best ones; ES and CMA-ES move one center brain toward the copies that scored better; PPO adjusts its one shared brain by backpropagation. Half the teams combine methods: evolution plus a PPO brain that joins the population, several PPO brains whose learning rates evolve, MAP-Elites driven by a CMA-ES search, novelty search (rewarding drivers that drive differently), and islands of different methods that swap their best drivers. Each team's best driver races in the recorded showcase replay.

      Test track

      The Grand Prix Oval: an oval with a long back straight (800 m) and a much shorter front straight (450 m), joined by turns that run in at an angle. 2.7 km, one lap, raced each way round in turn. Never used for training. Every 10 generations the champions race it to see whether the skill carries over.

      On this phone

      Everything runs in the browser, offline: the races in background workers (one per core), the learning on the page, and the results saved in the browser's own storage. Replays are re-simulated from the drivers' saved brains when you watch them, so each takes about 85 kB.

      Power

      How hard the phone works while training.

      Drivers per team

      More drivers means more races per generation: slower generations, steadier learning.

      Tracks per generation

      Each driver races this many different tracks before it is judged, and is scored on its average (per km, so long and short tracks count the same). More tracks: less luck in who gets picked and drivers that handle any track, but slower generations.

      Fitness score

      How drivers are scored after each race. Every team learns to score higher (the evolving teams breed from their best scorers; the PPO team gets these points as its rewards), so this decides what they learn to do.

      Changes count from the next batch of races. Earlier scores aren't recalculated.

      Storage

      Checking…

      Start over

      Deletes all training progress, stats and replays. The networks start again from random brains.