Struct CmaesCanonicalConfig 

pub struct CmaesCanonicalConfig {

    pub population_size: usize,
    pub max_generations: usize,
    pub seed: u64,
    pub c1: Option<f64>,
    pub c_mu: Option<f64>,
    pub c_sigma: Option<f64>,
    pub d_sigma: Option<f64>,
    pub parallel_eval: bool,
    pub verbosity: u8,
    pub ipop_restarts: usize,
    pub ipop_increase_factor: f64,
    pub bipop_restarts: usize,
    pub total_evals_budget: usize,
    pub use_subrun_budgeting: bool,
    pub alpha_mu: Option<f64>,
    pub hsig_threshold_factor: Option<f64>,
    pub bipop_small_population_factor: Option<f64>,
    pub bipop_small_budget_factor: Option<f64>,
    pub bipop_large_budget_factor: Option<f64>,
    pub bipop_large_pop_increase_factor: Option<f64>,
    pub max_bound_iterations: Option<usize>,
    pub eig_precision_threshold: Option<f64>,
    pub min_eig_value: Option<f64>,
    pub matrix_op_threshold: Option<f64>,
    pub stagnation_limit: Option<usize>,
    pub min_sigma: Option<f64>,
}

Configuration for canonical CMA-ES with evolution paths and rank updates.

This struct contains all parameters that control the behavior of the CMA-ES algorithm. Most parameters have sensible defaults and can be left as None to use automatic values.

Basic Parameters

The most important parameters for typical usage are:

• population_size: Controls exploration vs exploitation trade-off
• max_generations: Maximum number of generations to run
• parallel_eval: Enable parallel function evaluation
• verbosity: Control output level

Advanced Parameters

Advanced users can fine-tune the algorithm behavior using learning rates, restart strategies, and numerical precision settings.

Example

use cmaes_lbfgsb::cmaes::CmaesCanonicalConfig;
 
// Basic configuration
let config = CmaesCanonicalConfig {
    population_size: 20,
    max_generations: 1000,
    parallel_eval: true,
    verbosity: 1,
    ..Default::default()
};
 
// Advanced configuration with restarts
let advanced_config = CmaesCanonicalConfig {
    population_size: 50,
    max_generations: 500,
    ipop_restarts: 3,
    total_evals_budget: 100000,
    use_subrun_budgeting: true,
    ..Default::default()
};

Fields

population_size: usize

Population size (number of candidate solutions per generation).

Default: 0 (automatic: 4 + 3⌊ln(n)⌋ where n is problem dimension)

Typical range: 10-100 for most problems

Larger values: Better global exploration, slower convergence, more robust Smaller values: Faster convergence, risk of premature convergence

Guidelines:

• Easy problems (unimodal): Use smaller populations (10-20)
• Difficult problems (multimodal): Use larger populations (50-100+)
• High-dimensional problems: Consider automatic sizing

max_generations: usize

Maximum number of generations per run.

Default: 500

Guidelines:

• Simple problems: 100-500 generations usually sufficient
• Complex problems: 1000-5000+ generations may be needed
• Use with total_evals_budget for better control

If sub-run budgeting is disabled, each run uses this value for generations. Otherwise, sub-runs can get smaller or larger generation budgets.

seed: u64

Random seed for reproducible results.

Default: 42

Set to different values to get different random runs, or keep the same for reproducible experiments.

c1: Option<f64>

Learning rate for rank-one update of covariance matrix.

Default: None (automatic: 2/((n+1.3)² + μ_eff))

Typical range: 0.0001 - 0.1

Controls how quickly the algorithm adapts to the search direction. Smaller values = more conservative adaptation.

c_mu: Option<f64>

Learning rate for rank-μ update of covariance matrix.

Default: None (automatic: depends on μ_eff and problem dimension)

Typical range: 0.001 - 1.0

Controls how much the population covariance influences the search distribution. Must be balanced with c1 to ensure proper covariance matrix updates.

c_sigma: Option<f64>

Learning rate for cumulation path for step-size control.

Default: None (automatic: (μ_eff + 2)/(n + μ_eff + 5))

Typical range: 0.1 - 1.0

Controls the step-size adaptation speed. Larger values lead to faster step-size changes but may cause instability.

d_sigma: Option<f64>

Damping parameter for step-size update.

Default: None (automatic: 1 + 2max(0, √((μ_eff-1)/(n+1)) - 1))

Typical range: 1.0 - 10.0

Controls the damping of step-size updates. Larger values = more conservative step-size changes, which can improve stability but slow adaptation.

parallel_eval: bool

Enable parallel evaluation of candidate solutions.

Default: false

When true, the population is evaluated in parallel using Rayon. Recommended for expensive objective functions. Disable for very fast functions where parallelization overhead exceeds benefits.

verbosity: u8

Verbosity level for progress output.

Levels:

• 0: Silent (no output)
• 1: Basic progress (every 10 generations)
• 2: Detailed debug information

Default: 0

ipop_restarts: usize

Number of IPOP (Increasing Population) restarts.

Default: 0 (no IPOP restarts)

Typical range: 0-5 restarts

IPOP restarts the algorithm with increasing population sizes when it gets stuck. Each restart doubles the population size by default. Effective for multimodal problems but increases computational cost.

Note: BIPOP overrides IPOP if both are > 0.

ipop_increase_factor: f64

Factor by which population size is multiplied each IPOP restart.

Default: 2.0

Typical range: 1.5 - 3.0

Controls how aggressively the population size grows with each restart. Larger factors provide better exploration but increase cost exponentially.

bipop_restarts: usize

Number of BIPOP (Bi-Population) restarts.

Default: 0 (no BIPOP restarts)

Typical range: 0-10 restarts

BIPOP alternates between small and large population runs. More sophisticated than IPOP and often more effective for difficult multimodal problems.

Note: BIPOP overrides IPOP if both are > 0.

total_evals_budget: usize

Total function-evaluations budget across all runs.

Default: 0 (no budget limit)

Typical values: 10,000 - 1,000,000 depending on problem complexity

When combined with use_subrun_budgeting, this budget is intelligently allocated across multiple restart runs.

use_subrun_budgeting: bool

Enable advanced sub-run budgeting logic for IPOP/BIPOP.

Default: false

When true, the total_evals_budget is strategically allocated across restart runs rather than running each to completion. This often provides better results within a fixed computational budget.

alpha_mu: Option<f64>

Alpha parameter used in c_mu calculation.

Default: Some(2.0)

Typical range: 1.0 - 4.0

Advanced parameter that influences the balance between rank-1 and rank-μ updates of the covariance matrix. Rarely needs adjustment.

hsig_threshold_factor: Option<f64>

Threshold factor for the evolution path test in step-size control.

Default: Some(1.4)

Typical range: 1.0 - 2.0

Controls when to halt the cumulation of the evolution path based on its length. Affects the balance between exploration and exploitation.

bipop_small_population_factor: Option<f64>

Factor for small population size calculation in BIPOP.

Default: Some(0.5)

Typical range: 0.1 - 0.8

In BIPOP, determines the size of “small” population runs relative to the baseline population size. Smaller values = more focused local search.

bipop_small_budget_factor: Option<f64>

Budget allocation factor for small BIPOP runs.

Default: Some(1.0)

Typical range: 0.5 - 2.0

Controls how much of the evaluation budget is allocated to small population runs in BIPOP. Values > 1.0 give more budget to exploitation phases.

bipop_large_budget_factor: Option<f64>

Budget allocation factor for large BIPOP runs.

Default: Some(3.0)

Typical range: 1.0 - 5.0

Controls how much of the evaluation budget is allocated to large population runs in BIPOP. Values > 1.0 give more budget to exploration phases.

bipop_large_pop_increase_factor: Option<f64>

Factor for large population size increase in BIPOP.

Default: Some(2.0)

Typical range: 1.5 - 3.0

Controls how the large population size grows with each BIPOP restart. Similar to ipop_increase_factor but for BIPOP large runs.

max_bound_iterations: Option<usize>

Maximum number of iterations for bounds mirroring.

Default: Some(8)

Typical range: 5 - 20

When candidates violate bounds, they are “mirrored” back into the feasible region. This parameter limits how many mirror operations are performed before clamping to bounds.

eig_precision_threshold: Option<f64>

Numerical precision threshold for eigendecomposition convergence.

Default: Some(1e-15)

Typical range: 1e-20 to 1e-10

Controls the precision of the eigendecomposition of the covariance matrix. Smaller values = higher precision but potentially slower computation.

min_eig_value: Option<f64>

Minimum threshold for covariance matrix eigenvalues.

Default: Some(1e-15)

Typical range: 1e-20 to 1e-10

Prevents numerical issues by ensuring eigenvalues don’t become too small. Smaller values allow more aggressive adaptation but risk numerical instability.

matrix_op_threshold: Option<f64>

Minimum threshold for matrix operations.

Default: Some(1e-20)

Typical range: 1e-25 to 1e-15

General numerical threshold for matrix computations to prevent underflow and maintain numerical stability.

stagnation_limit: Option<usize>

Maximum stagnation generations before termination.

Default: Some(200)

Typical range: 50 - 1000

Algorithm terminates if no improvement is seen for this many consecutive generations. Prevents infinite runs on problems where the optimum has been reached within tolerance.

min_sigma: Option<f64>

Minimum sigma value to prevent numerical issues.

Default: Some(1e-8)

Typical range: 1e-12 to 1e-6

Prevents the step-size from becoming so small that progress stops due to numerical precision limits. Should be much smaller than expected parameter scales.

Trait Implementations

impl Default for CmaesCanonicalConfig

fn default() -> CmaesCanonicalConfig

Returns the “default value” for a type.