Stand H4

About the Project

This project evolves neural networks with NEAT (NeuroEvolution of Augmenting Topologies) to classify images even under adversarial alterations. Instead of training weights via backpropagation, the system uses evolutionary operators such as mutation, crossover, and speciation to discover both topology and weights. Where the main goals are: 1) Achieve high accuracy on clean MNIST images. 2) Maintain performance under adversarial inputs (e.g., FGSM). 3) Provide transparent visualization of evolved architectures and species dynamics. The core methods include: Representation (Genome): Node genes (input/bias/hidden/output) and connection genes (weighted, enabled/disabled) with innovation numbers for historical tracking. • Phenotype (Network): Built from a genome via topological sort; forward pass uses tanh for hidden layers; outputs are either linear or tanh-controlled by configuration (Set parameters button). • Fitness: Weighted accuracy on clean and adversarial subsets. • Speciation: Genomes are clustered by compatibility distance. The Compatibility Threshold controls how easily new species form. • Evolution Loop (per generation): 1) Evaluate fitness for all genomes. 2) Group into species. 3) Elitism + Hall-of-Fame injection to preserve top solutions. 4) Reproduction within species (crossover + mutations: add node/connection, toggle, weight alteration). 5) Replace population and track metrics (avg/best fitness, species sizes).