A cool 3D interactive demo of neuroevolution with supermario

A Mario-style platformer where a population of neural networks learns to play by evolution alone. No dependencies, no build step: open index.html (or run ./serve.ps1 and visit http://localhost:8080).

The evolution — level 1-2 stays stuck for ~75 generations, then a lucky mutation breaks through and finishers start to appear.

File Role

js/game.js Headless platformer: seeded procedural levels (3 themes), physics, goombas, spikes, pits, ? blocks, coins, flag

js/core.js Seeded RNG + Genome (feed-forward net 43-14-8-4 with per-connection on/off mask)

js/evolution.js Population loop: evaluate → rank → elites → tournament selection → crossover → mutation (weights and topology)

js/render.js Game renderer, "AI vision" overlay, on-screen gamepad that mirrors the output neurons

js/viz3d.js Interactive 3D network (orbit / zoom / hover), live activations, signal pulses, mutation flashes

js/main.js Fixed 60 Hz clock, UI wiring, charts, play-vs-AI mode

How the pieces connect

Game → network: the 8×5 tiles around Mario (solid / enemy / spike / coin) + on-ground, vertical and horizontal speed = 43 inputs. The dashed grid on the game screen is exactly the input layer in 3D.

Network → game: 4 outputs (◀ ▶ JUMP RUN) press the virtual gamepad.

Fitness: distance + coins + stomps, plus a large bonus for finishing quickly.

Evolution: top 8 % survive unchanged, the rest are bred from the top 40 % (crossover + Gaussian mutation + connection toggling), 4 % fresh random brains. Mutation rises automatically when progress stagnates.

Controls

Start/Pause (Space), Next gen (fast-forward to the end of the generation), Reset, speed 1×–MAX, level, population, mutation rate.

Click a bar in Population to pin that agent's brain in 3D; follow leader releases it.

Play & race the AI tab: arrows/WASD + Space (jump), Shift/Z (run), R restart — you race the best evolved champion as a ghost.

3D: drag to orbit, scroll to zoom, double-click to reset, hover a node for its value, slider hides weak connections.

Typical results: level 1 is usually solved within ~10–30 generations; level 2 shows a long plateau followed by a breakthrough; level 3 is a real challenge.

Handy for sharing a specific moment (and how the screenshots above were made): index.html?level=1&seed=3&warmup=85&frames=200&pause=1 — pick a level and seed, pre-evolve 85 generations instantly, run 200 frames into the next one, and pause. Add &only=game|brain|stats to show a single panel.