NEATGD teaches itself to play Geometry Dash levels using NEAT (NeuroEvolution of Augmenting Topologies); neural networks that evolve their own structure and weights from scratch.
Every genome is judged by actually playing the level: each one takes over your player for a fast-forwarded attempt, its network reading the terrain ahead and deciding when to jump, the exact same engine a human plays against, so what a genome scores in training is what it replays. Children inherit their parent's proven inputs up to just before where the parent died, then explore from that frontier, so progress is never lost to a bad mutation. The furthest, cleanest runs breed the next generation until something learns to beat the level.
- Open a level and hit pause.
- Click the auto button in the top left corner of the pause menu.
- Tune the training settings (each has an info button explaining it) and press Train.
- Watch the population evolve. When training finishes, the best genome plays the level back at normal speed.
Click the button again at any time to cancel.
- NEAT improves through generations, so a moderate population (~150-500) with a high generation cap usually learns faster than a huge one.
- Turn on Hide graphics to spend your hardware on physics instead of rendering. It lets you push the Speed value higher.
- this shi sucks because its my first ever mod and i hardly know C++🙏🙏🙏
- considering it IS my first mod im extremely proud of it given ive only learned C++ for about a month now
- if you have any suggestions or feedback hit me up on discord @itzar
Made by Itzar.