Evolutionary algorithms
Posted on Thu 28 May 2015 in Notes
Terms:
- Evolutionary Algorithm (EA)
- Population: Collection of string of symbols. In case of fooling an image classifier, the population is the list of images used throughout the test.
- Chromosomes: Each string (or image) in the population is called a chromosome.
- Initial population: This is our starting point. All chromosomes in the initial population are randomly (or heuristically) generated.
-
Fitness: Each chromosome has a fitness value attached to it. Fitness is determined by our fitness function. For [## Terms:
-
Evolutionary Algorithm (EA)
- Population: Collection of string of symbols. In case of fooling an image classifier, the population is the list of images used throughout the test.
- Chromosomes: Each string (or image) in the population is called a chromosome.
- Initial population: This is our starting point. All chromosomes in the initial population are randomly (or heuristically) generated.
- Fitness: Each chromosome has a fitness value attached to it. Fitness is determined by our fitness function. For](https://en.wikipedia.org/wiki/Travelling_salesman_problem "Traveling Salesman problem on Wikipedia") it is the total length travelled for a particular route (or chromosome). When fooling image classifiers, the fitness is the confidence the classifier deems a chromosome is part of a particular class.
- Evaluation: Evaluating a chromosome is simply determining its fitness value.
- Selection: The EA chooses the chromosomes with highest fitness following some rule (top 100, or everything above a threshold)
- Parent: A parent is simply a chromosome that has been selected. From parents we generate offsprings.
- Offspring: The EA will generate new chromosome from the selections. This can be done in several ways:
- Crossover / recombination: generating new chromosomes by recombining two or more parents
- Mutation: generating offsprings by modifying a single parent, typically in some random way.