Asteroids Learning (Demo)
Program that learns to avoid Asteroids by machine learning. The operation is almost the same as Flappy Learning with 16 sensors of distance all around the ship.
In This Fork
- ReLU activation function for better learning (however, the pre-trained robots are unaccustomed to it, so give them about 40 generations to show ReLU's improvement)
- Performance optimizations on Neuroevolution.js (Typed Arrays, loop optimizations, reduced variable declarations, etc.)
- Changed population from 50 to 200
- "Wall" distance sensor fixed
- localStorage save (between 10 generations)
- Added an array per ship to serve as "memory" and a way to the output manipulate this (and worked well)
- Save best score network as "hero"
- Trained network sample included
Extra
Notes
- Tested with Google Chrome and Mozilla Firefox