Reinforcement Learning Experiments (CarRacing-v2)

Learning-driven PPO/SAC experiments on CarRacing-v2 that emphasize clean pipelines, structured logging, and reproducible evaluation for control tasks.

Highlights

• PPO and SAC baselines with consistent seeding and environment wrappers.
• Training diagnostics: reward curves, episode length tracking, and rollout stats.
• Evaluation harness with saved checkpoints and deterministic test runs.

Approach

• Environment setup: Frame stacking, action smoothing, and observation normalization; wrappers for reproducibility and cleanup.
• Training loop: Config-driven runs, periodic evaluation episodes, and artifact saving for later comparison.
• Analysis: Simple notebooks for metric inspection and hyperparameter sweeps; attention to reset conditions and failure modes.

Artifacts

• Code and runs: github.com/ememchijioke/carracing-rl-experiments
• Logs/checkpoints for comparing policies across seeds and parameter sets.