white100big/LunarLander-v3-DQN

🚀 LunarLander-v3 DQN Autonomous Landing Agent

This repository contains a fully trained Double DQN with Dueling Architecture agent for the LunarLander-v3 environment in Gymnasium.

• Developer / Creator: white100big
• Algorithm: Double Dueling Deep Q-Network (DQN)
• Environment: Gymnasium LunarLander-v3
• Target Solved Score: 200.0+
• Evaluation Score: 271.25 (100% Landing Success Rate over test evaluations)

---

📈 1,000 Episodes Training Performance

Result: Successfully solved LunarLander-v3 with a 100-episode moving average reward of 259.09 (exceeding the standard 200.0 threshold) and a 100% landing success rate upon evaluation.

📊 Training Specifications

Parameter	Value
Total Episodes	1,000
Epsilon Schedule	1.0 (100%) → 0.05 (5%) Exponential Decay
Discount Factor ($\gamma$)	0.99
Replay Buffer Size	100,000
Batch Size	64
Loss Function	Smooth L1 (Huber Loss)
Target Network Update	Soft Update ($ au = 0.005$)
State Dimension	8
Action Space	Discrete(4)

---

🛠️ How to Use & Evaluate

1. Requirements

pip install torch "gymnasium[box2d]" swig

2. Download from Hugging Face

from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="white100big/LunarLander-v3-DQN",
    filename="best_lunarlander_dqn.pth"
)
print("Downloaded to:", model_path)

3. Run Autonomous Landing Simulation

import gymnasium as gym
from dqn_agent import DQNAgent

env = gym.make("LunarLander-v3", render_mode="human")
agent = DQNAgent(state_size=8, action_size=4)
agent.load("best_lunarlander_dqn.pth")

state, _ = env.reset()
done = False
total_reward = 0

while not done:
    action = agent.act(state, eps=0.0) # Pure exploitation
    state, reward, terminated, truncated, _ = env.step(action)
    done = terminated or truncated
    total_reward += reward

print(f"Final Landing Score: {total_reward:.2f}")
env.close()