How joshuadam/neat-javascript works
Project type and purpose
• NEAT-JavaScript is an open-source library implementing the NEAT (Neuroevolution of Augmenting Topologies) algorithm in JavaScript.
• Its purpose is to provide a reusable, extensible engine for evolving neural networks (weights and topology) in both Node.js and browser environments.
File structure
• Top‐level:
• dist/ — compiled bundle for browser.
• src/ — source code organized by responsibility.
• webpack.config.js — build configuration.
• src subdirectories:
• activationfunction/ — strategy objects for neural activation (Sigmoid, ReLU, Tanh, etc.).
• weightinitialization/ — strategies for initial weight assignment.
• core/
– genome/ — defines Genome class, subpackages for ConnectionGene, NodeGene, and supporting genetic-encoding data structures.
– population/ — Population and Species classes to manage speciation and generation lifecycle.
• algorithm/ — the NEAT orchestrator (speciation, crossover, mutation loop).
• config/ — Config object for global parameters (mutation rates, population size).
• fitnessfunction/ — plugin/factory for user‐defined fitness evaluations (e.g., XOR).
• util/ — helpers: GenomeBuilder, StaticManager, and trackers (InnovationTracker, NodeTracker, etc.).
• package.json & README.md describe usage, examples, and build.
README for additional insights
• Confirms core components: speciation, crossover, structural mutation.
• Describes usage in Node.js and browser; links to demos.
Main components & relationships
• Config Module → holds algorithm parameters, passed to Algorithm.
• Algorithm Engine → central controller:
• initializes Population from GenomeBuilder + Config
• for each generation: evaluates fitness via FitnessFunction plugins, applies Speciation (Species class), Crossover & Mutation (connection and node mutations)
• Core Data Structures:
• Genome ← composed of NodeGenes + ConnectionGenes
• GeneticEncoding metadata to persist innovation numbers
• Utility Trackers ← global singletons capturing innovation count, node IDs, species IDs
• Strategy Modules (ActivationFunction, WeightInitialization) ← injected into Network evaluation
• Dist Bundle ← exposes public API (Algorithm, Config, GenomeBuilder, fitness templates)
Architectural patterns & design principles
• Strategy pattern for activation & weight initialization.
• Factory for GenomeBuilder and FitnessFunction.
• Singleton trackers for global innovation management.
• Modular layering: Core data models → Algorithm orchestration → User plugins → I/O (Node.js or browser bundle).
Components
• Population & Species (subcomponent under Algorithm)
• ActivationFunction & WeightInitialization strategies (parallel modules)
• FitnessFunction plugin (user‐provided)
• Utility Trackers (InnovationTracker, NodeTracker)
• Dist Bundle / Public API layer
Detail
• Subcomponents, especially the trackers and genetic encoding classes.
• Extension points: fitnessfunction, activationfunction, weightinitialization.
• Don’t omit the build step (webpack) and the dual runtime (Node/browser).