Entry 1
Sun Nov 15, 2020
I began with https://github.com/alenaksu/neatjs, a javascript implementation of Kenneth Stanley's NEAT algorithm. I studied this codebase and began a rewrite in a branch of this repo. I made several changes to the core classes as I saw fit and wrote a few unit tests to make sure things were working. But then I got mired in some difficult bugs that introduced into some of the more complex methods, and I wasn't even sure if they were working in the first place. I timeboxed my efforts and eventually abandoned that branch after looking again a bit closer at https://github.com/wagenaartje/neataptic. This repo appeared to be laid out more sensibly, was better documented, and a few good example usages existed, such as https://github.com/wagenaartje/target-seeking-ai.
I started another branch and set up a basic implementation of a Game class that ran a simple simulation of a population created by neataptic. Every Player would spawn in the middle of a canvas (x:0,y:0) and then activate its network with the following inputs:
class Player {
step() {
const inputs = [
health.position.x,
health.position.y,
health.amount,
this.position.x,
this.position.y,
];
const [velocityX, velocityY] = this.network.activate(inputs);Where one health would be spawned randomly on the canvas somewhere. When a player's position
got close enough to the position of the health, that player's health would incease. Each step of
the simulation every player's health would decrement by one, or if a player moved out-of-bounds
their health would immediately be set to zero, effectively killing them. The goal here being
to evolve players that seek the health, basic target acquisition.
Before attempting to evolve, I would mutate the original population 100 times and then just run the simulation and render the paths of 100 players to seed with some variety. The following screenshots depict those initial experiments. You can see that with only 100 mutations of a simple network with 5 inputs and 2 outputs it produced a somewhat interesting assortment of movement behaviors:
Then I set about evolving the population, whereby each players fitness, or score, would be its health minus its distance to the health. Thus those players who got closer would be rewarded and those who got close enough would be rewarded more (by acquisition of the health).
With only 1000 rounds of evolution this usually yielded a decent champion that would move closer to the target health, as depicted in the following screenshots:
