.NET implementation and additions: Radiate.NET
Coming from Evolutionary Radiation.
Evolutionary radiation is a rapid increase in the number of species with a common ancestor, characterized by great ecological and morphological diversity - Pascal Neige.
By separating the the evolutionary process from the object being evolved, users can evolve any defined structure. The algorithm follows an evolutionary process through speciation which allows structures to optimize within their own niche.
Radiate exposes three traits to the user which must be implemented (full simple implementation below):
- Genome
Genome wraps the structure to be evolved and makes the user implement two necessary functions and an optional one. Distance and crossover must be implemented but base is optional (depending on how the user chooses to fill the population). - Environment
Environment represents the evolutionary environment for the genome, which means it can contain simple statistics for the population's evolution, or parameters for crossover and distance. Internally it is wrapped in a mutable thread-safe pointer so it is not intended to be shared for each genome, but rather exist only once per population. Environment requires no implementations of its one function, however depending on the use case environment exposes a function called reset which is intended to 'reset' the environment. - Problem
Problem is what gives a genome its fitness score. It requires two implemented functions: empty and solve. Empty is required and should return a base problem (think new()). Solve takes a genome and returns that genome's fitness score, so this is where the analysis of the current state of the genome occurs.
Radiate also comes with one model already built in and another one available for use on crates.io. Those being radiate_matrix_tree and NEAT, radiate_matrix_tree is available on crates.io and NEAT comes prepackaged with radiate.
Also known as Neuroevolution of Augmented Topologies, is the algorithm described by Kenneth O. Stanley in this paper. This NEAT implementation also includes a backpropagation function which operates much like traditional neural networks which propagate the input error back through the network to adjust the weights. In pair with the evolution engine, this can produce very nice and quick results. NEAT lets the use define how the network will be constructed, whether that be in a traditional neural network fashion where layers are stacked next to each other or with evolutionary topologies as explained in the paper. This means NEAT can be used in an evolutionary sense, through forward propagation and back propagation, or any combination of the two. There are examples of both in /examples. more color on Neat in radiate/src/models/
Radiate also supports off-machine training where you can set up a problem to solve on one machine, then send the parameters to it from another through Radiate Web.
The population is pretty easy to set up assuming the all traits have been implemented. The population is a higher abstraction to keep track of variables used during evolution but not needed within an epoch - things like the problem, solution, printing to the screen etc. A new population is filled initially with default settings:
pub fn new() -> Self {
Population {
// define the number of members to participate in evolution and be injected into the current generation
size: 100,
// determine if the species should be aiming for a specific number of species by adjusting the distance threshold
dynamic_distance: false,
// debug_progress is only used to print out some information from each generation
// to the console during training to get a glimpse of what is going on
debug_progress: false,
// create a new config to help the speciation of the population
config: Config::new(),
// create a new empty generation to be passed down through the population
curr_gen: Generation::<T, E>::new(),
// keep track of fitness score stagnation through the population
stagnation: Stagnant::new(0, Vec::new()),
// Arc<Problem> so the problem can be sent between threads safely without duplicating the problem,
// if the problem gets duplicated every time, a supervised learning problem with a lot of data could take up a large amount of memory
solve: Arc::new(P::empty()),
// create new solver settings which will hold the specific settings for the defined solver
// that will allow the structure to evolve through generations
environment: Arc::new(RwLock::new(E::default())),
// determine which genomes will live on and pass down to the next generation
survivor_criteria: SurvivalCriteria::Fittest,
// determine how to pick parents to reproduce
parental_criteria: ParentalCriteria::BiasedRandom
}
}A quick and easy example of implementing all the needed traits and running the genetic engine to generate a string print "hello world!". There are examples of how to run a Neat neural network in radiate/src/models/. To run this:
git clone https://github.com/pkalivas/radiate.git
cd radiate_legacy
cargo build --verbose && cargo run --bin helloworldOn my computer (Windows 10, x64-based, i7-7700 @ 4.20GHz, 32GB RAM) this finishes in less than half a second.
extern crate radiate;
extern crate rand;
use std::error::Error;
use std::time::Instant;
use std::sync::{Arc, RwLock};
use rand::Rng;
use radiate::prelude::*;
fn main() -> Result<(), Box<dyn Error>> {
let thread_time = Instant::now();
let (top, _) = Population::<Hello, HelloEnv, World>::new()
.size(100)
.populate_base()
.dynamic_distance(true)
.stagnation(10, vec![Genocide::KillWorst(0.9)])
.configure(Config {
inbreed_rate: 0.001,
crossover_rate: 0.75,
distance: 0.5,
species_target: 5
})
.run(|model, fit, num| {
println!("Generation: {} score: {:.3?}\t{:?}", num, fit, model.as_string());
(fit - 12.0).abs() < 0.1 || num == 500
})?;
println!("\nTime in millis: {}, solution: {:?}", thread_time.elapsed().as_millis(), top.as_string());
Ok(())
}Now create the problem which holds the target and actually scores the solvers. Note that the target data isn't being copied for each solver.
pub struct World { target: Vec<char> }
impl World {
pub fn new() -> Self {
World {
target: vec!['h', 'e', 'l', 'l', 'o', ' ', 'w', 'o', 'r', 'l', 'd', '!']
}
}
}
impl Problem<Hello> for World {
fn empty() -> Self { World::new() }
fn solve(&self, model: &mut Hello) -> f32 {
let mut total = 0.0;
for (index, letter) in self.target.iter().enumerate() {
if letter == &model.data[index] {
total += 1.0;
}
}
total
}
}Now define an environment to hold global data for crossover and distance, things like record/stat keeping, crossover probabilities, really anything that is needed globally is held in this.
#[derive(Debug, Clone)]
pub struct HelloEnv {
pub alph: Vec<char>,
}
impl HelloEnv {
pub fn new() -> Self {
HelloEnv {
alph: vec!['!', ' ', 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z'], // now i know my abcs..
}
}
}
/// implement Environment and Default for the HelloEnv, Environment is there in case you want the environment to be dynamic
impl Envionment for HelloEnv {}
impl Default for HelloEnv {
fn default() -> Self {
Self::new()
}
}Finally, define a solver. This is type 'Genome' in the evolutionary process and makes up the population of each generation.
#[derive(Debug, Clone, PartialEq)]
pub struct Hello {
pub data: Vec<char>
}
impl Hello {
pub fn new(alph: &[char]) -> Self {
let mut r = rand::thread_rng();
Hello {
data: (0..12)
.map(|_| alph[r.gen_range(0..alph.len())])
.collect()
}
}
pub fn as_string(&self) -> String {
self.data
.iter()
.map(|x| String::from(x.to_string()))
.collect::<Vec<_>>()
.join("")
}
}
/// implement genome for Hello
impl Genome<Hello, HelloEnv> for Hello {
// the first parent is always going to be the most fit parent
fn crossover(parent_one: &Hello, parent_two: &Hello, _env: &Arc<RwLock<HelloEnv>>, crossover_rate: f32) -> Option<Hello> {
let params = env.read().unwrap();
let mut r = rand::thread_rng();
let mut new_data = Vec::new();
if r.gen::<f32>() < crossover_rate {
for (one, two) in parent_one.data.iter().zip(parent_two.data.iter()) {
if one != two {
new_data.push(*one);
} else {
new_data.push(*two);
}
}
} else {
new_data = parent_one.data.clone();
let swap_index = r.gen_range(0..new_data.len());
new_data[swap_index] = params.alph[r.gen_range(0..params.alph.len())];
}
Some(Hello { data: new_data })
}
fn distance(one: &Hello, two: &Hello, _: &Arc<RwLock<HelloEnv>>) -> f32 {
let mut total = 0_f32;
for (i, j) in one.data.iter().zip(two.data.iter()) {
if i == j {
total += 1_f32;
}
}
one.data.len() as f32 / total
}
fn base(env: &mut HelloEnv) -> Hello {
Hello::new(&env.alph)
}
}The result looks something like this when running on the command line:
Generation: 100 score: 8.000 "!eulozworlde"
Generation: 101 score: 8.000 "!eulozworlde"
Generation: 102 score: 8.000 "!eulozworlde"
Generation: 103 score: 8.000 "!eulozworlde"
Generation: 104 score: 8.000 "!eulozworlde"
Generation: 105 score: 9.000 "heulozworlde"
Generation: 106 score: 9.000 "heulozworlde"
Generation: 107 score: 9.000 "heulozworlde"
Generation: 108 score: 9.000 "heulozworlde"
Generation: 109 score: 9.000 "heulozworlde"
Generation: 110 score: 9.000 "heulozworlde"
Generation: 111 score: 9.000 "heulozworlde"
Generation: 112 score: 10.000 "heulo worlde"
Generation: 113 score: 10.000 "heulo worlde"
Generation: 114 score: 10.000 "heulo worlde"
Generation: 115 score: 10.000 "heulo worlde"
Generation: 116 score: 10.000 "heulo worlde"
Generation: 117 score: 10.000 "heulo worlde"
Generation: 118 score: 10.000 "heulo worlde"
Generation: 119 score: 10.000 "heulo worlde"
Generation: 120 score: 11.000 "hello worlde"
Generation: 121 score: 11.000 "hello worlde"
Generation: 122 score: 11.000 "hello worlde"
Generation: 123 score: 11.000 "hello worlde"
Generation: 124 score: 11.000 "hello worlde"
Generation: 125 score: 11.000 "hello worlde"
Generation: 126 score: 12.000 "hello world!"
Time in millis: 349, solution: "hello world!"Right now there are four examples, just run "cargo run --bin (desired example name)" to run any of them
- xor-neat
- xor-neat-backprop
- lstm-neat
- helloworld
The initial generation in the population can be created in four different ways depending on the user's use case. The examples show different ways of using them.
- populate_gen - Give the population an already constructed Generation struct.
- populate_base - Create a generation of Genomes from the Genome's base function.
- populate_vec - Take a vec and populate the generation from the Genomes in the vec.
- populate_clone - Given a single Genome, clone it
sizetimes and create a generation from the clones.
Because the engine is meant to evolve Genomes through speciation, the Config struct is meant to hold parameters for the speciation of the population, adjusting these will change the way the Genomes are split up within the population and thus drive the discovery of new Genomes through crossover and mutation.
During evolution it can be common for either the population or specific species to become stagnant or stuck at a certain point in the problem space. To break out of this, population allows the user to define a number of stagnant generations until a 'genocide' will occur. These genocide options can be found in genocide.rs and are simply ways to clean the population to give the genomes an opportunity to breathe and evolve down a new path in the problem space.
pub enum Genocide {
KeepTop(usize),
KillWorst(f32),
KillRandom(f32),
KillOldestSpecies(usize)
}This is definitely an area which can be improved in the algorithm.
1.5.57 - Major improvements to the Dense/DensePool layers. Before the improvement the benchmark took about 1.5 minutes to run. With the improvements it finishes in about 1.5 seconds.
1.1.55 - Removed unsafe code and fixed memory leak in evtree. Refactored Evtree to be generic, moving neural network logic to be separate. Cleaned up send/sync impls.
1.1.52 - Added recurrent neurons for NEAT. Note - this is also only viable for evolution (I will focus on implementing backprop for recurrent neurons and GRU layers next - I'm working on some other projects that require recurrent evolution neurons as of now). This can be configured in the NeatEnvironment settings where the % change of adding a recurrent neuron can be added. 0.0 would mean no recurrent neurons are added, where 1.0 would mean every new neuron is recurrent. Example in radiate/src/models/.
1.1.5 - Added minimal support for GRU layer (Gated Recurrent Unit). GRU is only viable for evolution, NOT backprop, yet. Your program will panic! if a network with a gru layer is used during backprop, if a memory cell is needed for backprop purposes use the LSTM option for now. Also, cleaned up some code and optimized certain points to run a bit quicker. Evtree has been moved to it's own separate crate called radiate_matrix_tree and is available on crates.io.
1.1.3 - Adding support for Radiate Web so training Radiate can be done on a different machine.
1.1.2 - For forward and backward passes of NEAT, gated propagation in LSTM layers is now run in parallel which cuts training times in half. Changed the readme to be a full implementation of the engine which is a little more helpful for setting everything up. Added another readme file to radiate/src/models/ which gives examples of setting up a NEAT neural network.
1.1.1 - Fixed dumb bug in NEAT which was causing a error in backprop.
1.0.9 - As of 1/10/2020 all versions after 1.0.9 require the nightly toolchain Added serialization and deserialization to NEAT model through serde integration - serializing trait objects requires nightly crates for now.