Skip to content

Algorithm Configuration

Master detailed configuration options for each evolutionary algorithm.

Overview

Each evolutionary algorithm in EvoToolkit has its own configuration class with specific parameters. This tutorial covers all configuration options and how to tune them for your use case.

EvoEngineer Configuration

EvoEngineer is the primary LLM-driven evolutionary algorithm.

from evotoolkit.evo_method.evoengineer import EvoEngineerConfig

config = EvoEngineerConfig(
    interface=interface,
    output_path='./results',
    running_llm=llm_api,

    # Evolution parameters
    max_generations=20,      # Maximum number of generations
    pop_size=10,             # Population size
    max_sample_nums=15,      # Maximum samples per generation

    # Parallel control
    num_samplers=4,          # Number of parallel samplers
    num_evaluators=4,        # Number of parallel evaluators

    # Logging
    verbose=True             # Show verbose logging
)

Key Parameters

Evolution Parameters: - max_generations - Number of evolutionary generations to run - pop_size - Number of solutions to maintain in population - max_sample_nums - Maximum new solutions to sample per generation

Parallel Execution: - num_samplers - Parallel LLM sampling workers - num_evaluators - Parallel evaluation workers

Logging: - verbose - Enable detailed progress logging

Important Note

LLM temperature and other sampling parameters are set when creating HttpsApi, NOT in algorithm configuration.

from evotoolkit.tools.llm import HttpsApi

# LLM configuration happens here
llm_api = HttpsApi(
    api_key="your-key",
    model="claude-3-5-sonnet-20241022",
    temperature=0.7,  # LLM temperature
    max_tokens=4096
)

# Algorithm config does NOT include temperature
config = EvoEngineerConfig(
    running_llm=llm_api,  # Pass configured LLM
    # ... other params
)

FunSearch Configuration

FunSearch uses an island model for continuous evolution.

from evotoolkit.evo_method.funsearch import FunSearchConfig

config = FunSearchConfig(
    interface=interface,
    output_path='./results',
    running_llm=llm_api,

    # Sampling parameters
    max_sample_nums=30,           # Maximum number of samples
    programs_per_prompt=2,        # Programs generated per prompt

    # Island model
    num_islands=4,                # Number of parallel evolution islands
    max_population_size=1000,     # Maximum population size per island

    # Parallel control
    num_samplers=5,               # Number of parallel samplers
    num_evaluators=5,             # Number of parallel evaluators

    # Logging
    verbose=True
)

Key Parameters

Sampling: - max_sample_nums - Total samples to generate - programs_per_prompt - Solutions per LLM call

Island Model: - num_islands - Independent evolution islands (increases diversity) - max_population_size - Maximum solutions per island

Note: FunSearch does NOT use max_generations. It evolves continuously based on the island model until max_sample_nums is reached.

When to Use FunSearch

  • When you want continuous evolution without fixed generations
  • For exploring diverse solution spaces
  • When you have computational resources for large populations

EoH Configuration

EoH (Evolution of Heuristics) provides explicit control over genetic operators.

from evotoolkit.evo_method.eoh import EoHConfig

config = EoHConfig(
    interface=interface,
    output_path='./results',
    running_llm=llm_api,

    # Evolution parameters
    max_generations=10,       # Maximum number of generations
    pop_size=5,               # Population size
    max_sample_nums=20,       # Maximum samples per generation
    selection_num=2,          # Number of parents for crossover

    # Operator control
    use_e2_operator=True,     # Use E2 operator (crossover)
    use_m1_operator=True,     # Use M1 operator (mutation)
    use_m2_operator=True,     # Use M2 operator (second mutation)

    # Parallel control
    num_samplers=5,           # Number of parallel samplers
    num_evaluators=5,         # Number of parallel evaluators

    # Logging
    verbose=True
)

Key Parameters

Evolution: - max_generations - Number of generations - pop_size - Population size (typically smaller than EvoEngineer) - max_sample_nums - Samples per generation - selection_num - Parents selected for crossover

Genetic Operators: - use_e2_operator - Enable/disable crossover operator - use_m1_operator - Enable/disable first mutation operator - use_m2_operator - Enable/disable second mutation operator

When to Use EoH

  • When you want explicit control over genetic operators
  • For research comparing different operator combinations
  • When traditional EA concepts are important

Tuning Guidelines

Population Size

Small (5-10): - Pros: Faster generations, lower cost - Cons: Less diversity, may converge too quickly - Best for: Simple problems, limited resources

Medium (10-20): - Pros: Good balance of speed and diversity - Cons: None major - Best for: Most problems (recommended default)

Large (20+): - Pros: Maximum diversity, thorough exploration - Cons: Slower, higher cost - Best for: Complex problems, research

Parallel Execution

config = EvoEngineerConfig(
    # ... other params
    num_samplers=4,      # Parallel LLM calls
    num_evaluators=4,    # Parallel evaluations
)

Guidelines: - num_samplers: Set based on LLM API rate limits - num_evaluators: Set based on CPU/GPU availability - Start conservatively (2-4) and increase if resources allow

Example Configurations:

# Conservative (low resources)
num_samplers=2
num_evaluators=2

# Balanced (moderate resources)
num_samplers=4
num_evaluators=4

# Aggressive (high resources)
num_samplers=8
num_evaluators=8

Generation Count

Few Generations (5-10): - Quick experiments - Simple problems - Rapid prototyping

Medium Generations (10-20): - Most problems - Balanced exploration - Recommended default

Many Generations (20+): - Complex problems - Research studies - Final optimization runs

Configuration Presets

Quick Experimentation

config = EvoEngineerConfig(
    interface=interface,
    output_path='./results',
    running_llm=llm_api,
    max_generations=5,
    pop_size=5,
    max_sample_nums=10,
    num_samplers=2,
    num_evaluators=2,
    verbose=True
)

Use for: Testing, debugging, rapid iteration

Balanced Performance

config = EvoEngineerConfig(
    interface=interface,
    output_path='./results',
    running_llm=llm_api,
    max_generations=15,
    pop_size=10,
    max_sample_nums=20,
    num_samplers=4,
    num_evaluators=4,
    verbose=True
)

Use for: Most production use cases

config = EvoEngineerConfig(
    interface=interface,
    output_path='./results',
    running_llm=llm_api,
    max_generations=30,
    pop_size=15,
    max_sample_nums=30,
    num_samplers=6,
    num_evaluators=6,
    verbose=True
)

Use for: Research, benchmarking, final runs

Next Steps

Resources