lookarooka/LunarLander-v3-DQN

🚀 LunarLander-v3 Dueling DQN Agent

A trained Dueling Double Deep Q-Network (DQN) agent capable of smooth, stylish, and autonomous landing in the Gymnasium LunarLander-v3 environment with an average reward score of 200+ PTS.

---

🌟 Model Architecture & Hyperparameters

• Algorithm: Dueling Double DQN (Smooth L1 / Huber Loss)
• Input State Space: 8-dimensional continuous vector [x, y, vx, vy, angle, angular_vel, left_leg_contact, right_leg_contact]
• Action Space: Discrete(4) [0: No-op, 1: Left RCS, 2: Main Thruster, 3: Right RCS]
• Network:
  • Feature Extractor: Linear(8, 128) -> ReLU -> Linear(128, 128) -> ReLU
  • Value Stream $V(s)$: Linear(128, 64) -> ReLU -> Linear(64, 1)
  • Advantage Stream $A(s, a)$: Linear(128, 64) -> ReLU -> Linear(64, 4)
• Hyperparameters:
  • Learning Rate: 5e-4 (Adam)
  • Discount Factor $\gamma$: 0.99
  • Soft Target Update $\tau$: 0.005
  • Replay Buffer Size: 100,000
  • Batch Size: 64
  • Epsilon Schedule: Linear decay from 1.0 (100% exploration) to 0.05 (5% exploration)

---

🚀 How to Load and Run the Model

import gymnasium as gym
import torch
from dqn_agent import DQNAgent

# 1. Initialize environment & agent
env = gym.make("LunarLander-v3", render_mode="human")
agent = DQNAgent(state_dim=8, action_dim=4)

# 2. Load weights
agent.load("dqn_lunar_lander_best.pt")

# 3. Test pilot landing (Epsilon = 0)
state, info = env.reset()
total_reward = 0.0
done = False

while not done:
    action, _ = agent.act(state, evaluate=True)
    state, reward, terminated, truncated, info = env.step(action)
    done = terminated or truncated
    total_reward += reward

print(f"Landing Finished! Total Score: {total_reward:.1f} PTS")
env.close()

---

🛰️ Interactive Cyber-Space Web Dashboard

This model was trained with the LUNAR-LANDER-DQN 60FPS real-time web mission control platform.