CarRacing

Deep Q-learning on Gymnasium's CarRacing-v2 with discrete actions, learning directly from pixels.

• Model (DQN.py): a LeNet-style CNN (two conv + pool layers, three fully connected layers) that maps a 96×96 RGB frame to one Q-value per action, plus a replay memory.
• Training (main.py): DQN with a soft-updated target network (τ = 0.005), exponential ε-decay (1.0 → 0.05), AdamW (lr 1e-4) and batch size 256. It trains for 700 episodes on a GPU (50 on CPU), saves the policy and target weights as .pth files, and plots episode results.

LeNet like Structure.png and Figure_1.png are training plots. logs/best_model.zip is from an earlier PPO attempt that did not train correctly; the DQN agent is the one that learns.

Run

pip install "gymnasium[box2d]" torch matplotlib tqdm
python main.py

Status: work in progress (2023).