Configuration

The Config class provides a centralized way to configure all aspects of the NEAT algorithm. It manages parameters for network structure, evolution, speciation, and mutation with sensible defaults that can be overridden as needed.

Constructor

new Config(configObj = {})

Creates a new configuration object with default values that can be overridden by the provided configuration object.

Parameter	Type	Description
configObj	Object	Optional object containing configuration parameters to override defaults

// Create a configuration with default values
const config = new Config();

// Create a configuration with custom values
const customConfig = new Config({
  inputSize: 4,
  outputSize: 2,
  activationFunction: 'ReLU',
  populationSize: 200,
  allowRecurrentConnections: false
});

// Set values explicitly
customConfig.generations = 100;

Configuration Parameters

Network Structure

Parameter	Type	Default	Description
inputSize	Number	2	Number of input nodes in the network
outputSize	Number	1	Number of output nodes in the network
activationFunction	String/Object	'Sigmoid'	Activation function for the nodes. Can be specified as a string name from available functions or as a custom function object
weightInitialization	Object	{ type: 'Random', params: [-1, 1] }	Strategy for initializing connection weights. Can be specified as an object with type and parameters
minWeight	Number	-4.0	Minimum allowed weight value for connections
maxWeight	Number	4.0	Maximum allowed weight value for connections
allowRecurrentConnections	Boolean	true	Whether recurrent connections are allowed in the network
recurrentConnectionRate	Number	1.0	Rate for forming recurrent connections when they are allowed
useBias	Boolean	true	Enables or disables the use of bias nodes in the network
bias	Number	1.0	The value for bias inputs (typically 1.0)
useBiasWeights	Boolean	true	When true, applies weights to bias connections (bias*weight); when false, adds bias directly (+bias)
connectBias	Boolean	true	When true, automatically connects the bias node to all output nodes during network construction
addBiasEverywhere	Boolean	false	When true, adds bias to every node's activation function (standard neural network approach)

Evolution Parameters

Parameter	Type	Default	Description
populationSize	Number	150	Size of the population
generations	Number	100	Maximum number of generations for evolution
fitnessFunction	String/Object	'XOR'	Function to evaluate genome fitness. Can be specified as a string name from available functions or as a custom function object
survivalRate	Number	0.2	Proportion of population that survives each generation
numOfElite	Number	10	Number of elite individuals to preserve unchanged in each generation
populationStagnationLimit	Number	20	Maximum number of generations without fitness improvement before population is considered stagnant
interspeciesMatingRate	Number	0.001	Rate of mating between individuals from different species
mutateOnlyProb	Number	0.25	Probability of producing offspring through mutation only (without crossover)

Speciation Parameters

Parameter	Type	Default	Description
c1	Number	1.0	Coefficient for excess genes in compatibility distance calculation
c2	Number	1.0	Coefficient for disjoint genes in compatibility distance calculation
c3	Number	0.4	Coefficient for weight differences in compatibility distance calculation
compatibilityThreshold	Number	3.0	Maximum compatibility distance for genomes to be considered part of the same species
dropOffAge	Number	15	Maximum age for a species before it is removed if no fitness improvement is observed

Mutation Parameters

Parameter	Type	Default	Description
mutationRate	Number	1.0	Overall probability of applying mutation to offspring
weightMutationRate	Number	0.8	Probability of mutating connection weights
addConnectionMutationRate	Number	0.05	Probability of adding a new connection
addNodeMutationRate	Number	0.03	Probability of adding a new node
perturbRange	Number	0.5	Range for weight perturbation when mutating weights
keepDisabledOnCrossOverRate	Number	0.75	Probability of inheriting disabled genes during crossover

Available Built-in Functions

Activation Functions

Name	Description
'Sigmoid'	Standard sigmoid function: f(x) = 1 / (1 + e^(-x))
'NEATSigmoid'	Modified sigmoid function as used in the original NEAT paper
'Tanh'	Hyperbolic tangent: f(x) = tanh(x)
'ReLU'	Rectified Linear Unit: f(x) = max(0, x)
'LeakyReLU'	Leaky Rectified Linear Unit: f(x) = x if x > 0, else αx
'Gaussian'	Gaussian function: f(x) = e^(-x²)

Weight Initializations

Name	Parameters	Description
'Random'	[min, max]	Initializes weights randomly within the specified range

Fitness Functions

Name	Description
'XOR'	Evaluates fitness based on the XOR problem (built-in test case)

Using Custom Functions

Custom Activation Function

// Create a custom activation function
class CustomActivation {
  constructor(param = 1.0) {
    this.param = param;
  }

  activate(x) {
    return Math.sin(x * this.param);
  }

  derivative(x) {
    return this.param * Math.cos(x * this.param);
  }
}

// Use the custom activation function in configuration
const config = new Config({
  activationFunction: new CustomActivation(0.5)
});

Custom Fitness Function

// Create a custom fitness function
class CartPoleFunction {
  constructor(simulationSteps = 500) {
    this.simulationSteps = simulationSteps;
  }

  calculateFitness(genome) {
    let fitness = 0;
    // Run cart-pole simulation with the genome
    // ...
    return fitness;
  }
}

// Use the custom fitness function in configuration
const config = new Config({
  inputSize: 4,  // Cart position, velocity, pole angle, pole velocity
  outputSize: 1, // Force direction
  fitnessFunction: new CartPoleFunction(1000)
});