This repository contains two javascript-based neuroevolution demos using the swirlnet library, which is based on NEAT. They attempt to solve two 2D physical balancing tasks. With a few commands, winning network behavior can be graphically demonstrated in the browser.
Install runtime dependencies:
$ npm install
Install web demo build dependencies:
# npm install -g browserify
To watch the pre-generated networks' performance in the browser:
$ cd src/doublePendulum
$ make
Then open up file:///.../src/doublePendulum/index.html in a browser window.
$ cd src/doublePole
$ make
Then open up file:///.../src/doublePole/index.html in a browser window.
To search for new solutions:
$ node demos/doublePole/solve.js
$ node demos/doublePendulum/solve.js
To watch a network you have found:
$ cd demos/doublePole
$ mkdir solutions
$ edit solutions/net-0.json # paste in the json from the above solver's output
$ ln -sf solutions/net-0.json latest-solution.json
$ make
Then open up file:///.../demos/doublePole/index.html in a browser window.
Of the two demos, the double pole experiment solves faster. Its code is also easier to comprehend.
The double pendulum experiement may find the partial solution of balancing already upright pendulums, but it is unlikely to swing up both pendulums and balance them. To fully solve that problem, the fitness function and testing scenarios need to be adjusted.
As networks become fitter on average they will tend to survive longer in challenges. This is why later generations take longer to test.
This software is licensed under the Apache License, Version 2.0
The authors of swirlnet claim no license over the networks, genomes and phenotypes evolved using this library.