About elite_selection_ratio and exploitation_ratio

[Ethan-han921]

Dose
 elite_selection_ratio  + exploitation_ratio + random = 1 ?

Can I set the

 elite_selection_ratio: 1.0
 exploitation_ratio: 1.0 

in my Config.yaml

Since I want to get the score be improved every iteration.

[codelion]

There are three Ratios.

exploration_ratio + exploitation_ratio + random = 1.0

From the code in database.py:979-990, the parent selection works as follows:

  if rand_val < self.config.exploration_ratio:
      # EXPLORATION: Sample from current island
  elif rand_val < self.config.exploration_ratio + self.config.exploitation_ratio:
      # EXPLOITATION: Sample from archive (elite programs)
  else:
      # RANDOM: Sample from any program (remaining probability)

The default values are:

• exploration_ratio: 0.2 (20%)
• exploitation_ratio: 0.7 (70%)
• Random: 0.1 (10%) - implicitly the remainder

About elite_selection_ratio:

elite_selection_ratio is different - it's not part of the parent selection probability. Instead, it determines what
percentage of programs are considered "elite" when selecting top programs for various operations (see database.py:1138).

We cannot set both exploitation_ratio: 1.0 and elite_selection_ratio: 1.0 because:

1. If exploitation_ratio: 1.0, then all parent selection would be from the archive (no exploration or random selection)
2. You'd need exploration_ratio: 0.0 to make the probabilities sum to 1.0
3. elite_selection_ratio: 1.0 is technically possible but means all programs are considered "elite"

The valid constraint is:

• exploration_ratio + exploitation_ratio ≤ 1.0
• The remainder is the random selection probability

For maximum exploitation, you could use:

database:
exploration_ratio: 0.0
exploitation_ratio: 1.0  # All parents from archive
elite_selection_ratio: 0.1  # Top 10% are elite (this is separate)

[Laxyny]

It looks like these parameters are used in two different parts of the algorithm. The parent selection process uses three probabilities: exploration_ratio, exploitation_ratio and an implicit "random" probability. The random probability is simply whatever is left over when you subtract exploration and exploitation from 1.0. In the default configuration, exploration is 0.2, exploitation is 0.7 and random is 0.1; these three probabilities add up to 1.

elite_selection_ratio is separate – it defines what fraction of your population is considered "elite" when building the archive or migrating programs, and it doesn’t participate in the parent selection probabilities.
"elite" when building the archive or migrating programs, and it doesn’t participate in the parent selection probabilities.

Because of that separation you can’t set both elite_selection_ratio and exploitation_ratio to 1.0. If you want exploitation to be 100 %, you must set exploration_ratio to 0.0 so that exploration + exploitation ≤ 1.0; then there’s no exploration and no random sampling. And an elite_selection_ratio of 1.0 would mark every program as elite, which defeats the purpose of keeping a top-performers archive.

For maximum exploitation while still keeping an elite archive you could do something like this:

database:
  exploration_ratio: 0.0
  exploitation_ratio: 1.0  # All parents from archive
  elite_selection_ratio: 0.1  # Top 10 % considered elite

Those values adhere to the default config and the selection logic in database.py. Hope that helps clarify how the ratios relate to each other!

[Ethan-han921]

It looks like these parameters are used in two different parts of the algorithm. The parent selection process uses three probabilities: exploration_ratio, exploitation_ratio and an implicit "random" probability. The random probability is simply whatever is left over when you subtract exploration and exploitation from 1.0. In the default configuration, exploration is 0.2, exploitation is 0.7 and random is 0.1; these three probabilities add up to 1.

elite_selection_ratio is separate – it defines what fraction of your population is considered "elite" when building the archive or migrating programs, and it doesn’t participate in the parent selection probabilities. "elite" when building the archive or migrating programs, and it doesn’t participate in the parent selection probabilities.

Because of that separation you can’t set both elite_selection_ratio and exploitation_ratio to 1.0. If you want exploitation to be 100 %, you must set exploration_ratio to 0.0 so that exploration + exploitation ≤ 1.0; then there’s no exploration and no random sampling. And an elite_selection_ratio of 1.0 would mark every program as elite, which defeats the purpose of keeping a top-performers archive.

For maximum exploitation while still keeping an elite archive you could do something like this:

database:
exploration_ratio: 0.0
exploitation_ratio: 1.0 # All parents from archive
elite_selection_ratio: 0.1 # Top 10 % considered elite
Those values adhere to the default config and the selection logic in database.py. Hope that helps clarify how the ratios relate to each other!

Thank you so much.

This is my config.yaml for Configuration for function minimization example

max_iterations: 10
checkpoint_interval: 5
log_level: "INFO"

# LLM configuration
llm:
  primary_model: "o4-mini"  
  primary_model_weight: 1.0
  api_base: "https://api.openai.com/v1"
  api_key: "sk-xxxxx"
  max_tokens: 4096

# Prompt configuration
prompt:

  include_artifacts: true
  system_message: "You are an expert programmer specializing in optimization algorithms. Your task is to improve a function minimization algorithm to find the global minimum of a complex function with many local minima. The function is f(x, y) = sin(x) * cos(y) + sin(x*y) + (x^2 + y^2)/20. Focus on improving the search_algorithm function to reliably find the global minimum, escaping local minima that might trap simple algorithms."
  num_top_programs: 1
  max_artifact_bytes: 4096
  use_template_stochasticity: true
  artifact_security_filter: true

# Database configuration
database:
  population_size:  1
  archive_size: 1
  num_islands: 1
  exploration_ratio: 0.0
  elite_selection_ratio: 0.1
  exploitation_ratio: 1.0


# Evaluator configuration
evaluator:
  timeout: 60
  cascade_evaluation: true
  enable_artifacts: true


# Evolution settings
diff_based_evolution: false

but the score is
0.681818
0.568182
0.590909

It likes not an stable increased trend.

[Ethan-han921]

There are three Ratios.

exploration_ratio + exploitation_ratio + random = 1.0

From the code in database.py:979-990, the parent selection works as follows:

  if rand_val < self.config.exploration_ratio:
      # EXPLORATION: Sample from current island
  elif rand_val < self.config.exploration_ratio + self.config.exploitation_ratio:
      # EXPLOITATION: Sample from archive (elite programs)
  else:
      # RANDOM: Sample from any program (remaining probability)

The default values are:

• exploration_ratio: 0.2 (20%)
• exploitation_ratio: 0.7 (70%)
• Random: 0.1 (10%) - implicitly the remainder

About elite_selection_ratio:

elite_selection_ratio is different - it's not part of the parent selection probability. Instead, it determines what percentage of programs are considered "elite" when selecting top programs for various operations (see database.py:1138).

We cannot set both exploitation_ratio: 1.0 and elite_selection_ratio: 1.0 because:

1. If exploitation_ratio: 1.0, then all parent selection would be from the archive (no exploration or random selection)
2. You'd need exploration_ratio: 0.0 to make the probabilities sum to 1.0
3. elite_selection_ratio: 1.0 is technically possible but means all programs are considered "elite"

The valid constraint is:

• exploration_ratio + exploitation_ratio ≤ 1.0
• The remainder is the random selection probability

For maximum exploitation, you could use:

database:
exploration_ratio: 0.0
exploitation_ratio: 1.0  # All parents from archive
elite_selection_ratio: 0.1  # Top 10% are elite (this is separate)

Thank you so much. It's so clear