Eevee - Generic Neuroevolution

This project is a WIP!

Nothing really works very well. There's a lot of useful code, and topology search / genome evolution can be done, but it's slow, inefficient, and often fails completely. Expect frequent changes.

Eevee doesn't work on Windows. This is because our default RNG seeding assumes that /dev/urandom exists.

Overview

Eevee is a library for leveraging the NEAT algorithm to train genomes encoding neural network behavior. Specifically, it aims to generalize the algorithm such that it may be applied to different domains, and maybe in the future applications that don't implement neural networks at all.

Ok... but why is it named after a Pokemon

I like to name lots of my projects after Pokemon. I called this one Eevee because, like generic NeuroEvolution, Eevee can evolve in a number of different ways - all of which fill their own niche, are good at some things, and not so good at others. Also because docs.rs/eevee wasn't occupied.

Try it and see

The core iteration loop is that, given a scenario which implements some mechanism by which a genome may be scored with a fitness, Eevee will try mutate, reproduce, and cull genomes to optimize for that fitness to increase. There exist some experiments aound this in the examples folder.

Building on it

It's written in rust, and uses nightly versions - mostly for incomplete features. Later when I add CUDA support, it will even more rely on nightly.

I use criterion for benchmarking, it's recommended that if you run benches, you have gnuplot on your system. You can use ./cmp-bench <bench> [branch:-] to compare a benchmark across two branches, which produces a nice report.

I use flamegraph for profiling, it's required that if you run benches with profiling, you have perf on your system. You can use ./profile <bench> to run benchmarks on a pared-down version of any benchmark, and ./cmp-profile <bench> [branch:-] to compare a profiling across two branches.

For both of those, I use toml-cli + jq to get a list of benchmarks.

$ toml get Cargo.toml . | jq '.bench | map(.name)[]' -r

I use tarpaulin for getting test coverage.

Other things

Thanks to smol-rs/fastrand, I stole the core of their WyHash rng implementation.

Thanks to TLmaK0/rustneat, This project turned me on to learning about CTRNN's and also I stole the CTRNN matmul code