Below are my stream of consciousness notes on Chapter 2. Overall not a lot of new meat to chew in this chapter, but I appreciate the time taken to ensure the basic concepts are established before moving into the deep end.
Ch 2
Populations should use data structures that implement Enumerable protocol.
Using List as default will be simplest form of population.
-
Initialize population
- Random list of
chromosomes(i.e., possible solutions) - Initialization function should be generic (i.e., not care structure used by chromosome)
- Random list of
-
Evaluate population
- Apply the fitness function (i.e., sort population according to fitness criteria)
- Fitness will be defined unique to each problem
population -> fitness.() -> sorted_population
-
Select parents
- Elitism selection pairs strongest chromosomes as parents
- This is the point my background gets in the way of the metaphor. In biology, a population is a collection of organisms. Each organism can have one or many chromosomes. Each chromosome can have one or many genes. Each gene can have one or many values. Unclear if this is just the simplest implementation of a genetic algorithm and in future chapters or more advanced implementations the additional nested structure is utilized more consistent with the biology metaphor.
population -> transform_into_parent_pairs.() -> population_ready_for_crossover
-
Creating children
- Apply the crossover function to parents
parents -> crossover.() -> childrenwhere size(children) == size(initial_pop)
-
Mutation
- Prevents premature convergence
population -> mutate.() -> x_menwheremutateonly affects a small % of population
-
Termination criteria
- criteria are problem-specific
- simplest case is reaching desired value known ahead of time (max value in one-max e.g.)
-
Hyperparameters
- set prior to training algorithm
- includes pop size, mutation rate, etc
- use optional list of parameters in keyword list to allow rapidly adjusting these params
(I would have instinctively used module attributes, but this is more opaque and less
generic for a framework)
Using mix run <project_root>/scripts/<script>.exs is a pretty cool structure for evaluating examples using a lib.
I let it run for about 20k generations and the score didn’t seem like it was moving.