Neat-Ex v1.3.0 Neat.Options
Below are the options that can be configured the Neat evolution process.
| Paramater | Default | Description |
|---|---|---|
| :seed | Required | The seed ANN used to generate the initial population. This is used for specifying the input/output neurons that all ANNs should have, as well as any priliminary topology ANNs should start with. |
| :fitness_function | Required | The function used to evaluate the fitness of ANNs. |
| :single_fitness_function | Required | Should equal true or false. If true, fitness function is given a single neural network to evaluate, and it should return a positive fitness value (where greater numbers are better). The call looks like fitness_function(ann). If false, the fitness function is given the entire population to evaluate. The call looks like fitness_function([{species_rep1, [ann1, {ann2, old_fitness}, ann3, ...]}, ...]). Note, the species_rep should be left alone. Also, some ann’s in the population (such as ann2 in this example) will already have assigned fitnesses, and can either be re-evaluated, or left with the same fitness. All ann’s should be replaced by {ann, fitness} tuples, and the entire structure should be returned. Single-fitness functions are far easier to implement, but multi-fitness functions provide more flexibility. |
| :population_size | 150 | The number of ANNs in any given generation. |
| :target_species_number | 15 | The ideal number of species to exist. |
| :structure_diff_coefficient | 2.0 | Coefficient for how much structure differences matter in speciation. |
| :weight_diff_coefficient | 1.0 | Coefficient for how much weight differences matter in speciation. |
| :compatibility_threshold | 6.0 | Initial threshold used for speciation, changes quickly. |
| :compatibility_modifier | 0.3 | Value used to change :compatibility_threshold. |
| :dropoff_age | 15 | If a species does not progress for this many generations, it dies. |
| :survival_ratio | 0.2 | Ratio of the population that persists each generation. |
| :weight_range_min | -1.0 | Minimum for initial weight creation. |
| :weight_range_max | 1.0 | Maximum for initial weight creation. |
| :weight_mutation_power | 1/6 | Standard deviation for the random modifiers used to mutate weights. |
| :new_weight_chance | 0.1 | Ratio of weight mutations that get assigned totally new values. |
| :child_from_mutation_chance | 0.25 | Ratio of children that result from mutation (not mating). |
| :interspecies_mating_chance | 0.01 | Ratio of children that result from interspecies mating. |
| :new_node_chance | 0.03 | Ratio of mutations that add a new node (the rest mutate weights). |
| :new_link_chance | 0.05 | Ratio of mutations that add a new link (the rest mutate weights). |
| :recurrent_nodes | true | Whether or not recurrent nodes are allowed. |
Summary
Functions
Creates an new Neat.Options given a seed network (in which the initial population is based), the fitness_function, and a map of other (optional) paramaters
Functions
Creates an new Neat.Options given a seed network (in which the initial population is based), the fitness_function, and a map of other (optional) paramaters.
If :population_size is specified, and :target_species_number is not, it’s value is 1/10th of the :population_size.
Example Usage:
iex> Neat.Options.new(Ann.new([1, 2, 3], [4]), &fitness/1, [population_size: 20])