GEMLA/gemla/src/bracket/genetic_node.rs
2021-08-23 10:04:11 -07:00

528 lines
17 KiB
Rust

//! A trait used to interact with the internal state of nodes within the [`Bracket`]
//!
//! [`Bracket`]: crate::bracket::Bracket
use serde::{Deserialize, Serialize};
use std::fmt;
/// An enum used to control the state of a [`GeneticNode`]
///
/// [`GeneticNode`]: crate::bracket::genetic_node
#[derive(Clone, Debug, Serialize, Deserialize, Copy)]
#[serde(tag = "enumType", content = "enumContent")]
pub enum GeneticState {
/// The node and it's data have not finished initializing
Initialize,
/// The node is currently simulating a round against target data to determine the fitness of the population
Simulate,
/// The node is currently selecting members of the population that scored well and reducing the total population size
Score,
/// The node is currently mutating members of it's population and breeding new members
Mutate,
/// The node has finished processing for a given number of iterations
Finish,
}
/// A trait used to interact with the internal state of nodes within the [`Bracket`]
///
/// [`Bracket`]: crate::bracket::Bracket
pub trait GeneticNode {
/// Initializes a new instance of a [`GeneticState`].
///
/// # Examples
///
/// ```
/// # use gemla::bracket::genetic_node::GeneticNode;
/// #
/// struct Node {
/// pub fit_score: f64,
/// }
///
/// impl GeneticNode for Node {
/// fn initialize() -> Result<Box<Self>, String> {
/// Ok(Box::new(Node {fit_score: 0.0}))
/// }
///
/// //...
/// #
/// # fn simulate(&mut self, iterations: u64) -> Result<(), String> {
/// # Ok(())
/// # }
/// #
/// # fn get_fit_score(&self) -> f64 {
/// # self.fit_score
/// # }
/// #
/// # fn calculate_scores_and_trim(&mut self) -> Result<(), String> {
/// # Ok(())
/// # }
/// #
/// # fn mutate(&mut self) -> Result<(), String> {
/// # Ok(())
/// # }
/// }
///
/// # fn main() -> Result<(), String> {
/// let node = Node::initialize()?;
/// assert_eq!(node.get_fit_score(), 0.0);
/// # Ok(())
/// # }
/// ```
fn initialize() -> Result<Box<Self>, String>;
/// Runs a simulation on the state object for the given number of `iterations` in order to guage it's fitness.
/// This will be called for every node in a bracket before evaluating it's fitness against other nodes.
///
/// #Examples
///
/// ```
/// # use gemla::bracket::genetic_node::GeneticNode;
/// #
/// struct Model {
/// pub fit_score: f64,
/// //...
/// }
///
/// struct Node {
/// pub models: Vec<Model>,
/// //...
/// }
///
/// impl Model {
/// fn fit(&mut self, epochs: u64) -> Result<(), String> {
/// //...
/// # self.fit_score += epochs as f64;
/// # Ok(())
/// }
/// }
///
/// impl GeneticNode for Node {
/// # fn initialize() -> Result<Box<Self>, String> {
/// # Ok(Box::new(Node {models: vec![Model {fit_score: 0.0}]}))
/// # }
/// #
/// //...
///
/// fn simulate(&mut self, iterations: u64) -> Result<(), String> {
/// for m in self.models.iter_mut()
/// {
/// m.fit(iterations)?;
/// }
/// Ok(())
/// }
///
/// //...
///
/// # fn get_fit_score(&self) -> f64 {
/// # self.models.iter().max_by(|m1, m2| m1.fit_score.partial_cmp(&m2.fit_score).unwrap()).unwrap().fit_score
/// # }
/// #
/// # fn calculate_scores_and_trim(&mut self) -> Result<(), String> {
/// # Ok(())
/// # }
/// #
/// # fn mutate(&mut self) -> Result<(), String> {
/// # Ok(())
/// # }
/// }
///
/// # fn main() -> Result<(), String> {
/// let mut node = Node::initialize()?;
/// node.simulate(5)?;
/// assert_eq!(node.get_fit_score(), 5.0);
/// # Ok(())
/// # }
/// ```
fn simulate(&mut self, iterations: u64) -> Result<(), String>;
/// Returns a fit score associated with the nodes performance.
/// This will be used by a bracket in order to determine the most successful child.
///
/// # Examples
/// ```
/// # use gemla::bracket::genetic_node::GeneticNode;
/// #
/// struct Model {
/// pub fit_score: f64,
/// //...
/// }
///
/// struct Node {
/// pub models: Vec<Model>,
/// //...
/// }
///
/// # impl Model {
/// # fn fit(&mut self, epochs: u64) -> Result<(), String> {
/// # //...
/// # self.fit_score += epochs as f64;
/// # Ok(())
/// # }
/// # }
///
/// impl GeneticNode for Node {
/// # fn initialize() -> Result<Box<Self>, String> {
/// # Ok(Box::new(Node {models: vec![Model {fit_score: 0.0}]}))
/// # }
/// #
/// # //...
/// #
/// # fn simulate(&mut self, iterations: u64) -> Result<(), String> {
/// # for m in self.models.iter_mut()
/// # {
/// # m.fit(iterations)?;
/// # }
/// # Ok(())
/// # }
/// #
/// //...
///
/// fn get_fit_score(&self) -> f64 {
/// self.models.iter().max_by(|m1, m2| m1.fit_score.partial_cmp(&m2.fit_score).unwrap()).unwrap().fit_score
/// }
///
/// //...
/// # fn calculate_scores_and_trim(&mut self) -> Result<(), String> {
/// # Ok(())
/// # }
/// #
/// # fn mutate(&mut self) -> Result<(), String> {
/// # Ok(())
/// # }
/// }
///
/// # fn main() -> Result<(), String> {
/// let mut node = Node::initialize()?;
/// node.simulate(5)?;
/// assert_eq!(node.get_fit_score(), 5.0);
/// # Ok(())
/// # }
/// ```
fn get_fit_score(&self) -> f64;
/// Used when scoring the nodes after simulating and should remove underperforming children.
///
/// # Examples
/// ```
/// # use gemla::bracket::genetic_node::GeneticNode;
/// #
/// struct Model {
/// pub fit_score: f64,
/// //...
/// }
///
/// struct Node {
/// pub models: Vec<Model>,
/// population_size: i64,
/// //...
/// }
///
/// # impl Model {
/// # fn fit(&mut self, epochs: u64) -> Result<(), String> {
/// # //...
/// # self.fit_score += epochs as f64;
/// # Ok(())
/// # }
/// # }
///
/// impl GeneticNode for Node {
/// # fn initialize() -> Result<Box<Self>, String> {
/// # Ok(Box::new(Node {
/// # models: vec![
/// # Model { fit_score: 0.0 },
/// # Model { fit_score: 1.0 },
/// # Model { fit_score: 2.0 },
/// # Model { fit_score: 3.0 },
/// # Model { fit_score: 4.0 },
/// # ],
/// # population_size: 5,
/// # }))
/// # }
/// #
/// # //...
/// #
/// # fn simulate(&mut self, iterations: u64) -> Result<(), String> {
/// # for m in self.models.iter_mut() {
/// # m.fit(iterations)?;
/// # }
/// # Ok(())
/// # }
/// #
/// //...
///
/// # fn get_fit_score(&self) -> f64 {
/// # self.models
/// # .iter()
/// # .max_by(|m1, m2| m1.fit_score.partial_cmp(&m2.fit_score).unwrap())
/// # .unwrap()
/// # .fit_score
/// # }
/// #
/// fn calculate_scores_and_trim(&mut self) -> Result<(), String> {
/// self.models.sort_by(|a, b| a.fit_score.partial_cmp(&b.fit_score).unwrap().reverse());
/// self.models.truncate(3);
/// Ok(())
/// }
///
/// //...
/// #
/// # fn mutate(&mut self) -> Result<(), String> {
/// # Ok(())
/// # }
/// }
///
/// # fn main() -> Result<(), String> {
/// let mut node = Node::initialize()?;
/// assert_eq!(node.models.len(), 5);
///
/// node.simulate(5)?;
/// node.calculate_scores_and_trim()?;
/// assert_eq!(node.models.len(), 3);
///
/// # assert_eq!(node.get_fit_score(), 9.0);
/// # Ok(())
/// # }
/// ```
fn calculate_scores_and_trim(&mut self) -> Result<(), String>;
/// Mutates members in a population and/or crossbreeds them to produce new offspring.
///
/// # Examples
/// ```
/// # use gemla::bracket::genetic_node::GeneticNode;
/// # use std::convert::TryInto;
/// #
/// struct Model {
/// pub fit_score: f64,
/// //...
/// }
///
/// struct Node {
/// pub models: Vec<Model>,
/// population_size: i64,
/// //...
/// }
///
/// # impl Model {
/// # fn fit(&mut self, epochs: u64) -> Result<(), String> {
/// # //...
/// # self.fit_score += epochs as f64;
/// # Ok(())
/// # }
/// # }
///
/// fn mutate_random_individuals(_models: &Vec<Model>) -> Model
/// {
/// //...
/// # Model {
/// # fit_score: 0.0
/// # }
/// }
///
/// impl GeneticNode for Node {
/// # fn initialize() -> Result<Box<Self>, String> {
/// # Ok(Box::new(Node {
/// # models: vec![
/// # Model { fit_score: 0.0 },
/// # Model { fit_score: 1.0 },
/// # Model { fit_score: 2.0 },
/// # Model { fit_score: 3.0 },
/// # Model { fit_score: 4.0 },
/// # ],
/// # population_size: 5,
/// # }))
/// # }
/// #
/// # fn simulate(&mut self, iterations: u64) -> Result<(), String> {
/// # for m in self.models.iter_mut() {
/// # m.fit(iterations)?;
/// # }
/// # Ok(())
/// # }
/// #
/// # fn get_fit_score(&self) -> f64 {
/// # self.models
/// # .iter()
/// # .max_by(|m1, m2| m1.fit_score.partial_cmp(&m2.fit_score).unwrap())
/// # .unwrap()
/// # .fit_score
/// # }
/// #
/// # fn calculate_scores_and_trim(&mut self) -> Result<(), String> {
/// # self.models.sort_by(|a, b| a.fit_score.partial_cmp(&b.fit_score).unwrap().reverse());
/// # self.models.truncate(3);
/// # Ok(())
/// # }
/// //...
///
/// fn mutate(&mut self) -> Result<(), String> {
/// loop {
/// if self.models.len() < self.population_size.try_into().unwrap()
/// {
/// self.models.push(mutate_random_individuals(&self.models))
/// }
/// else{
/// return Ok(());
/// }
/// }
/// }
/// }
///
/// # fn main() -> Result<(), String> {
/// let mut node = Node::initialize()?;
/// assert_eq!(node.models.len(), 5);
///
/// node.simulate(5)?;
/// node.calculate_scores_and_trim()?;
/// assert_eq!(node.models.len(), 3);
///
/// node.mutate()?;
/// assert_eq!(node.models.len(), 5);
///
/// # assert_eq!(node.get_fit_score(), 9.0);
/// # Ok(())
/// # }
/// ```
fn mutate(&mut self) -> Result<(), String>;
}
/// Used externally to wrap a node implementing the [`GeneticNode`] trait. Processes state transitions for the given node as
/// well as signal recovery. Transition states are given by [`GeneticState`]
#[derive(Serialize, Deserialize, Clone, Debug)]
pub struct GeneticNodeWrapper<T>
where
T: GeneticNode,
{
pub data: Option<T>,
state: GeneticState,
pub iteration: u32,
}
impl<T> GeneticNodeWrapper<T>
where
T: GeneticNode + fmt::Debug,
{
/// Initializes a wrapper around a GeneticNode. If the initialization is successful the internal state will be changed to
/// `GeneticState::Simulate` otherwise it will remain as `GeneticState::Initialize` and will attempt to be created in
/// [`process_node`](#method.process_node).
///
/// # Examples
/// ```
/// # use gemla::bracket::genetic_node::GeneticNode;
/// # use gemla::bracket::genetic_node::GeneticNodeWrapper;
/// # #[derive(Debug)]
/// struct Node {
/// # pub fit_score: f64,
/// //...
/// }
///
/// impl GeneticNode for Node {
/// //...
/// # fn initialize() -> Result<Box<Self>, String> {
/// # Ok(Box::new(Node {fit_score: 0.0}))
/// # }
/// #
/// #
/// # fn simulate(&mut self, iterations: u64) -> Result<(), String> {
/// # Ok(())
/// # }
/// #
/// # fn get_fit_score(&self) -> f64 {
/// # self.fit_score
/// # }
/// #
/// # fn calculate_scores_and_trim(&mut self) -> Result<(), String> {
/// # Ok(())
/// # }
/// #
/// # fn mutate(&mut self) -> Result<(), String> {
/// # Ok(())
/// # }
/// }
///
/// # fn main() -> Result<(), String> {
/// let mut wrapped_node = GeneticNodeWrapper::<Node>::new()?;
/// assert_eq!(wrapped_node.data.unwrap().get_fit_score(), 0.0);
/// # Ok(())
/// # }
/// ```
pub fn new() -> Result<Self, String> {
let mut node = GeneticNodeWrapper {
data: None,
state: GeneticState::Initialize,
iteration: 0,
};
let new_data = T::initialize()?;
node.data = Some(*new_data);
node.state = GeneticState::Simulate;
Ok(node)
}
/// Performs state transitions on the [`GeneticNode`] wrapped by the [`GeneticNodeWrapper`].
/// Will loop through the node training and scoring process for the given number of `iterations`.
///
/// ## Transitions
/// - `GeneticState::Initialize`: will attempt to call [`initialize`] on the node. When done successfully will change
/// the state to `GeneticState::Simulate`
/// - `GeneticState::Simulate`: Will call [`simulate`] with a number of iterations (not for `iterations`). Will change the state to `GeneticState::Score`
/// - `GeneticState::Score`: Will call [`calculate_scores_and_trim`] and when the number of `iterations` have been reached will change
/// state to `GeneticState::Finish`, otherwise it will change the state to `GeneticState::Mutate.
/// - `GeneticState::Mutate`: Will call [`mutate`] and will change the state to `GeneticState::Simulate.`
/// - `GeneticState::Finish`: Will finish processing the node and return.
///
/// [`initialize`]: crate::bracket::genetic_node::GeneticNode#tymethod.initialize
/// [`simulate`]: crate::bracket::genetic_node::GeneticNode#tymethod.simulate
/// [`calculate_scores_and_trim`]: crate::bracket::genetic_node::GeneticNode#tymethod.calculate_scores_and_trim
/// [`mutate`]: crate::bracket::genetic_node::GeneticNode#tymethod.mutate
pub fn process_node(&mut self, iterations: u32) -> Result<(), String> {
// Looping through each state transition until the number of iterations have been reached.
loop {
match (self.state, self.data.as_ref()) {
(GeneticState::Initialize, _) => {
self.iteration = 0;
let new_data =
T::initialize().map_err(|e| format!("Error initializing node: {}", e))?;
self.data = Some(*new_data);
self.state = GeneticState::Simulate;
}
(GeneticState::Simulate, Some(_)) => {
self.data
.as_mut()
.unwrap()
.simulate(5)
.map_err(|e| format!("Error simulating node: {}", e))?;
self.state = GeneticState::Score;
}
(GeneticState::Score, Some(_)) => {
self.data
.as_mut()
.unwrap()
.calculate_scores_and_trim()
.map_err(|e| format!("Error scoring and trimming node: {}", e))?;
self.state = if self.iteration == iterations {
GeneticState::Finish
} else {
GeneticState::Mutate
}
}
(GeneticState::Mutate, Some(_)) => {
self.data
.as_mut()
.unwrap()
.mutate()
.map_err(|e| format!("Error mutating node: {}", e))?;
self.state = GeneticState::Simulate;
}
(GeneticState::Finish, Some(_)) => {
break;
}
_ => return Err(format!("Error processing node {:?}", self.data)),
}
}
Ok(())
}
}