Evolution Strategies and Continuous Optimization
Evolution Strategies
Overview
Evolution strategies are evolutionary methods designed primarily for real-valued optimization. They often rely on Gaussian mutation and may adapt step sizes during the search process.
ES Terminology
Gaussian Mutation
A real-valued candidate is often mutated as , where is sampled from a standard normal distribution and controls the step size.
Basic ES Pseudocode
-
1
Step 1: Initialize a population of real-valued vectors and step sizes.
-
2
Step 2: Evaluate each vector using the objective function.
-
3
Step 3: Select parents.
-
4
Step 4: Mutate parents using Gaussian noise.
-
5
Step 5: Optionally adapt step sizes.
-
6
Step 6: Choose survivors by or selection.
-
7
Step 7: Repeat.
GA vs ES
Genetic Algorithm
- Often uses crossover heavily
- Common for discrete encodings
- Can also support real vectors
Evolution Strategy
- Usually emphasizes mutation over crossover
- Specialized for continuous parameters
- Often includes self-adaptation
What distribution is commonly used for mutation in evolution strategies?
Gaussian noise is the standard mutation mechanism in many ES methods.
Correct answer: Gaussian
What does the step size control?
It determines how far offspring move in parameter space.
Correct answer: The scale of mutation noise.
Which selection scheme allows parents to survive into the next generation?
selection includes both parents and offspring.
Correct answer:
Intuition
Small step sizes encourage fine-tuning; large step sizes encourage broader exploration.
Self-Adaptation Idea
In self-adaptive ES, strategy parameters such as are encoded in the individual and evolved along with solution variables, allowing the algorithm to adjust its own search behavior.