Search Configuration

Configure evolutionary search algorithms and database settings in SkyDiscover

Overview

The search section controls the evolutionary search algorithm, database backend, and context program selection strategy.

Basic Configuration

search:
  type: "topk"              # Search algorithm type
  num_context_programs: 4   # Number of examples in prompt
  output_dir: null          # Override output directory
  database:                 # Algorithm-specific settings
    db_path: null
    log_prompts: true

Search Types

SkyDiscover supports multiple search algorithms, each with different strategies for exploring the solution space.

Top-K Search

Simple best-first search that maintains the top-k programs by fitness.

search:
  type: "topk"
  num_context_programs: 4
  database:
    db_path: "outputs/database.db"
    log_prompts: true

db_path · str · default: "None"

Path to SQLite database for storing results

log_prompts · bool · default: "true"

Whether to log prompts to the database

AdaEvolve (Adaptive Multi-Island)

Adaptive evolutionary algorithm with multiple islands that adjust search intensity based on improvement history.

configs/adaevolve.yaml

search:
  type: "adaevolve"
  num_context_programs: 4
  database:
    # Population settings
    population_size: 20
    num_islands: 2
    
    # Adaptive search intensity
    decay: 0.9
    intensity_min: 0.15
    intensity_max: 0.5
    
    # Feature flags
    use_adaptive_search: true
    use_ucb_selection: true
    use_migration: true
    use_unified_archive: true
    
    # Migration
    migration_interval: 15
    migration_count: 5
    
    # Archive settings
    fitness_weight: 1.0
    novelty_weight: 0.0
    diversity_strategy: "code"  # code, metric, or hybrid
    
    # Dynamic islands
    use_dynamic_islands: true
    max_islands: 5
    spawn_productivity_threshold: 0.015
    spawn_cooldown_iterations: 30
    
    # Paradigm breakthrough
    use_paradigm_breakthrough: true
    paradigm_window_size: 10
    paradigm_improvement_threshold: 0.12
    paradigm_max_uses: 2
    paradigm_max_tried: 10
    paradigm_num_to_generate: 3

Defined in skydiscover/config.py:380-432

Population Settings

population_size · int · default: "20"

Number of programs per island

num_islands · int · default: "2"

Initial number of islands

archive_size · int · default: "100"

Maximum size of the unified archive

Adaptive Search

use_adaptive_search · bool · default: "true"

Enable adaptive search intensity based on improvement signal

decay · float · default: "0.9"

Exponential moving average weight (rho) for improvement signal

intensity_min · float · default: "0.15"

Minimum search intensity (exploitation mode)

intensity_max · float · default: "0.5"

Maximum search intensity (exploration mode)

fixed_intensity · float · default: "0.4"

Fixed intensity when use_adaptive_search=false

Selection & Migration

use_ucb_selection · bool · default: "true"

Use UCB (Upper Confidence Bound) for island selection

use_migration · bool · default: "true"

Enable inter-island migration

migration_interval · int · default: "15"

Migrate every N iterations

migration_count · int · default: "5"

Number of top programs to migrate

Archive & Diversity

use_unified_archive · bool · default: "true"

Use quality-diversity archive

fitness_weight · float · default: "1.0"

Weight for fitness rank in elite score

novelty_weight · float · default: "0.0"

Weight for novelty rank in elite score

diversity_strategy · str · default: "code"

Diversity metric: code, metric, or hybrid

k_neighbors · int · default: "5"

Number of neighbors for novelty calculation

Dynamic Islands

use_dynamic_islands · bool · default: "true"

Enable dynamic island spawning

max_islands · int · default: "5"

Maximum number of islands

spawn_productivity_threshold · float · default: "0.015"

Spawn new island if productivity drops below this

spawn_cooldown_iterations · int · default: "30"

Wait N iterations between island spawns

Paradigm Breakthrough

use_paradigm_breakthrough · bool · default: "true"

Generate new high-level strategies during stagnation

paradigm_window_size · int · default: "10"

Window for improvement rate calculation

paradigm_improvement_threshold · float · default: "0.12"

Trigger paradigm breakthrough below this improvement rate

paradigm_max_uses · int · default: "2"

Maximum uses per paradigm

paradigm_num_to_generate · int · default: "3"

Number of paradigms to generate per trigger

paradigm_max_tried · int · default: "10"

Maximum tried paradigms to track

OpenEvolve Native (MAP-Elites)

Quality-diversity search using MAP-Elites grids with island-based populations.

configs/openevolve_native.yaml

search:
  type: "openevolve_native"
  num_context_programs: 5
  database:
    num_islands: 5
    population_size: 40
    archive_size: 100
    
    # Selection strategies
    exploration_ratio: 0.2    # P(explore) - random from island
    exploitation_ratio: 0.7   # P(exploit) - archive elite
    # remaining 0.1 = P(random)
    
    elite_selection_ratio: 0.1
    
    # MAP-Elites features
    feature_dimensions: ["complexity", "diversity"]
    feature_bins: 10
    diversity_reference_size: 20
    
    # Migration
    migration_interval: 10
    migration_rate: 0.1
    
    random_seed: 42

Defined in skydiscover/config.py:435-450

num_islands · int · default: "5"

Number of islands in the population

population_size · int · default: "40"

Total population size across all islands

archive_size · int · default: "100"

Size of the MAP-Elites archive

exploration_ratio · float · default: "0.2"

Probability of exploration (random parent from current island)

exploitation_ratio · float · default: "0.7"

Probability of exploitation (elite from archive)

elite_selection_ratio · float · default: "0.1"

Fraction of context programs from top elites

feature_dimensions · List[str] · default: "[\"complexity\", \"diversity\"]"

Behavioral dimensions for MAP-Elites grid

feature_bins · int · default: "10"

Number of bins per feature dimension

GEPA Native

Guided Evolution for Program Adaptation with elite pool, epsilon-greedy selection, and LLM-mediated merge.

search:
  type: "gepa_native"
  num_context_programs: 4
  database:
    population_size: 40
    
    # Selection
    candidate_selection_strategy: "epsilon_greedy"  # epsilon_greedy, best, pareto
    epsilon: 0.1
    max_rejection_history: 20
    
    # Controller settings
    acceptance_gating: true
    use_merge: true
    merge_after_stagnation: 15
    max_merge_attempts: 10
    max_recent_failures: 5
    
    random_seed: 42

Defined in skydiscover/config.py:453-472

Beam Search

Beam search with diversity weighting.

search:
  type: "beam_search"
  num_context_programs: 4
  database:
    beam_width: 5
    beam_selection_strategy: "diversity_weighted"
    beam_diversity_weight: 0.3
    beam_temperature: 1.0
    beam_depth_penalty: 0.0

Defined in skydiscover/config.py:362-370

Best-of-N

Generate N candidates and select the best one.

search:
  type: "best_of_n"
  num_context_programs: 4
  database:
    best_of_n: 5

Defined in skydiscover/config.py:373-377

EvoX (Co-Evolution)

Label-guided co-evolutionary search.

configs/evox.yaml

search:
  type: "evox"
  database:
    database_file_path: null  # Auto-configured
    evaluation_file: null     # Auto-configured
    config_path: null         # Auto-configured
    auto_generate_variation_operators: true

Defined in skydiscover/config.py:339-359

SearchConfig Parameters

Defined in skydiscover/config.py:485-493

type · str · default: "topk"

Search algorithm: topk, adaevolve, openevolve_native, gepa_native, beam_search, best_of_n, or evox

num_context_programs · int · default: "4"

Number of example programs to include in generation prompts

output_dir · str · default: "None"

Override output directory. If None, auto-generates based on search type and timestamp

database · DatabaseConfig · default: "DatabaseConfig()"

Algorithm-specific database configuration

CLI Overrides

Override search type from command line:

# Use AdaEvolve
skydiscover-run program.py evaluator.py -s adaevolve

# Use OpenEvolve Native
skydiscover-run program.py evaluator.py -s openevolve_native

# Use beam search
skydiscover-run program.py evaluator.py -s beam_search

Choosing a Search Algorithm

Top-K

Best for: Quick experiments, simple optimization

Pros: Fast, simple, low overhead

Cons: Limited exploration

AdaEvolve

Best for: Complex optimization, long runs

Pros: Adaptive, robust, handles stagnation

Cons: More computational overhead

OpenEvolve Native

Best for: Quality-diversity, exploring trade-offs

Pros: Diverse solutions, explores full space

Cons: Requires good feature dimensions

GEPA

Best for: Merging diverse solutions

Pros: LLM-mediated combination, smart gating

Cons: Higher LLM token usage

Next Steps

LLM Configuration

Configure models for search

Evaluator Configuration

Set up program evaluation