Struct RegularizedEvolution 

pub struct RegularizedEvolution {
    pub population_size: usize,
    pub tournament_size: usize,
    /* private fields */
}

Regularized (aging) evolution NAS searcher.

Uses a cyclic population with tournament selection and single-step mutation.

Fields

population_size: usize

Target population size.

tournament_size: usize

Number of random members sampled per tournament.

Implementations

impl RegularizedEvolution

pub fn new( space: ArchSearchSpace, population_size: usize, tournament_size: usize, seed: u64, ) -> TrainResult<Self>

Create a new regularized evolution searcher.

Arguments

• space - Architecture search space.
• population_size - Target size of the cyclic population (≥ 2).
• tournament_size - Number of randomly drawn population members per tournament (≤ population_size).
• seed - RNG seed for reproducibility.

Errors

Returns TrainError::InvalidParameter when:

• population_size < 2
• tournament_size == 0 or tournament_size > population_size

pub fn ask(&mut self) -> TrainResult<Architecture>

Ask for the next architecture to evaluate.

• While the population has fewer than population_size members (warm-up phase), returns a freshly sampled random architecture.
• Once the population is full, performs tournament selection among tournament_size randomly chosen members and returns the winner mutated by one step.

pub fn tell(&mut self, arch: Architecture, score: f64)

Tell the result of evaluating an architecture.

Appends (arch, score) to the full history and pushes it into the population (at the back / newest position). When the population exceeds population_size, the oldest entry (front) is evicted.

pub fn best(&self) -> Option<&(Architecture, f64)>

Return a reference to the highest-scored (architecture, score) pair currently in the population, or None if the population is empty.

pub fn result(&self) -> Option<NasResult>

Produce a NasResult summarising the current search state.

Returns None if no evaluations have been recorded yet.