Arcade, simulation, and rendering

Arcade is a Python library for building and displaying 2D games. Here it provides the rendered window, event loop, and drawing API; it is not the neural-network or genetic-algorithm library.

Game owns one PlayZone per individual. At each simulation step it gathers all players' observations into a NumPy batch, uses PyTorch to predict their actions together, then applies the actions and advances every zone. TIME_OUT ends an epoch after the configured elapsed duration.

The rendered window draws one representative play zone and its neural network visualization; the remaining population continues to be simulated without individual rendering. Use python -m src.main for rendered training or python -m src.train for headless training with progress logs. Both run until interrupted unless --generations sets a limit.

The gameplay representative is a display choice, not a promise that the fitness-ranked elite is being shown. After fitness ranking, the first-ranked individual is checkpointed and used for validation; that checkpoint is written before validation runs.

Rendering uses a high-contrast dark court with distinct player/CPU paddle colors and a bright ball halo. Court markings, palette, and network-panel styling are draw-only; they do not feed into observations or change simulation state.

Rendered games generate short paddle-hit, wall-bounce, and scoring tones with Pyglet's synthesis API, so no audio files or new dependencies are needed. Only the displayed zone plays audio, with per-event cooldowns to keep rapid rallies from becoming noisy. Sound is enabled by default in src.main and src.tester; pass --no-sound to either command to disable it. The headless trainer is silent, and audio events do not affect physics or fitness.

After each generation, the best individual is evaluated in fresh games with varied starting positions and directions, duration, and paddle width. The headless evaluator and rendered suite share the scenario setup and result aggregation. Every scenario uses the same fixed simulation delta in either mode, and Python, NumPy, and PyTorch RNG states are isolated and restored before selection and reproduction. Its win and CPU-shutout rates are recorded alongside each run's settings, metrics, elite, and generation checkpoints. Reaching the configured validation target sets a persistent green indicator in the rendered runner; training does not stop. python -m src.tester loads the newest run's elite for a one-player, unbounded viewing game.