NEAT-AI: A Modern Neural Network Framework Integrating Evolutionary Algorithms and Gradient Optimization

NEAT-AI is a DenoJS/TypeScript-based neural network project that integrates modern techniques like error-guided discovery, MCMC mutation acceptance, and synthetic synapses on the foundation of the classic NEAT algorithm, supporting distributed training and lifelong learning. It retains NEAT's core idea (topology evolution) and balances development efficiency and performance through a hybrid architecture, making it suitable for DenoJS developers, neuroevolution researchers, and enterprises needing interpretable AI systems.

Neuroevolution is an important branch of AI that uses evolutionary algorithms to optimize neural network structures and weights. In 2002, Stanley and Miikkulainen proposed the NEAT algorithm, pioneering the direction of topology evolution. NEAT-AI is developed by the Australian stSoftware team, clearly distinguishing itself from the original NEAT (NEAT refers to the 2002 algorithm, while NEAT-AI is a new project). It retains the core contributions of classic NEAT: speciation (protecting innovative structures), structural mutation (gradually increasing complexity), and historical marking (tracking gene origins).

NEAT-AI adopts a three-layer architecture: 1. The TypeScript layer handles evolutionary logic (speciation, selection, mutation, reproduction); 2. The WebAssembly layer performs forward propagation and score calculation, providing high-performance activation functions; 3. The optional Rust layer provides GPU-accelerated error-guided structure discovery via FFI. Without the Rust extension, the discovery phase is skipped, and evolution runs entirely on WASM, enabling graceful degradation. This design balances development convenience and native performance.

NEAT-AI's core features include: 1. Error-guided discovery: Dynamically create new synapses by analyzing neuron activation and errors; the Rust extension accelerates this analysis with GPU, making structural changes directional (in contrast to traditional NEAT's random mutations); 2. Memory evolution: Lamarckian evolution that records the biases and weights of optimal individuals to fine-tune future generations; 3. Synthetic synapses: Temporarily add zero-weight synapses during training to increase density, then prune useless connections after training; 4. MCMC mutation acceptance: Use the Metropolis-Hastings criterion with adaptive temperature adjustment (target acceptance rate of 23.4%) to balance exploration and convergence.

NEAT-AI supports: 1. Distributed training: Island model with multi-node parallel training, where a central controller merges optimal individuals; 2. Lifelong learning: The same population continues training with new data to adapt to changing environments; 3. Scalable observation and transfer learning: Use UUIDs to identify features, no need to restart evolution; pre-trained organisms can export checkpoints to transfer to new tasks; 4. ONNX export: Trained models can be exported to ONNX format for integration into production ML pipelines.

NEAT-AI offers rich configurations: adaptive mutation rates (large organisms focus on weight modification), training data blurring (injecting noise to prevent overfitting), K-fold cross-validation, adaptive hyperparameters (each organism carries its own learning rate, etc.), CRISPR (injecting manual genes), and grafting algorithms (cross-island hybridization).

NEAT-AI is a modern attempt in the field of neuroevolution, integrating classic NEAT with modern ML technologies to balance development efficiency and performance. It is recommended for DenoJS developers to build neural networks; researchers to explore the combination of classic algorithms and modern technologies; and enterprises to use its end-to-end toolchain to develop interpretable and evolvable AI systems. The NEAT algorithm has remained viable for over two decades, and NEAT-AI injects new vitality into it.