Architecture

This section explains how AI-Pong evolves and evaluates its neural-network players. It is organized by concern:

• Genetic algorithm covers selection, crossover, and mutation.
• Neural network covers the player model, its activations, and the modular gameplay-explanation tooling.
• Simulation and rendering covers Arcade, batched inference, and audio.
• Animation pipeline covers regenerating the static diagrams and the source-driven gameplay trace/animation workflow.

Evolution loop

src.main creates an initial population of IndividualPlayer instances and passes it to GeneticAlgorithm in src/ga/ga_core.py. Each individual owns a neural network and one PlayZone. Every generation:

1. The population plays for the configured generation duration; each player is evaluated in its own simulation zone.
2. Fitness is calculated and the population is ranked. The fitness winner is checkpointed and tested on fresh validation scenarios (2 by default).
3. Selection and reproduction produce the next population.

Rendered training advances through this lifecycle in one Arcade window and event loop. Updates perform bounded gameplay, fitness, or validation work so the window can continue drawing and processing close events. Closing it cancels an in-progress validation and prevents another generation from starting. Headless training runs the same simulation synchronously without a window.

Population size and duration

The default population is 200. Training continues until interrupted unless --generations sets a limit.