Neuroevolution-Bots is a personal project that demonstrates neuroevolution in a browser environment using TensorFlow.js, Neataptic (for neural nets) and HTML5 Canvas (for graphics). I tried to create a scaled down 2D version of the popular Gym's Humanoid-v2 environment using Planck.js, a JavaScript rewrite of Box2D.
The bots in the simulation have simple brains (neural nets) that control their movements. Input to the brains: bot's head position in the world space and angles of bot's leg and knee joints; output of the brains: rotation forces to bot's leg and knee joints. After every iteration, a fitness function chooses the most fit bots to the next iteration (fitness score of bot: how much the bot traveled to the right + whether bot fell down or not).
Two variations of neuroevolution are available:
- NEAT (network architecture and weights are updated) (using Neataptic.js)
- Vanilla Neuroevolution (only network weights are updated) (using Tensorflow.js)
TRY THE DEMO: here
References and inspirations:
- Uber's blogpost on Deep Neuroevolution
- Evolving AI Lab posts
- Roboschool by OpenAI
- Evolution Simulator by Aditya Thebe
- Tweet by hardmaru
License: MIT