lunarlander-v3-dqn-physical-ai

🛸 LunarLander-v3 Double Dueling DQN (Physical AI)

This repository contains a trained Double Dueling Deep Q-Network (DQN) model designed for flawless, stylish landings in the Gymnasium LunarLander-v3 environment.

🎯 Model & Training Details

• Environment: LunarLander-v3
• Algorithm: Double Dueling DQN (Deep Q-Network)
• State Dimension: 8 (Coordinates, velocities, angle, angular velocity, leg touch sensors)
• Action Space: Discrete(4) - [0: Idle, 1: Fire Left, 2: Fire Main, 3: Fire Right]
• Epsilon Schedule: 100% ($1.0$) $\rightarrow$ 5% ($0.05$)
• Loss Function: Huber Loss (Smooth L1 Loss) with Soft Target Update ($\tau = 10^{-3}$)
• Reward Shaping: Stylish upright orientation + smooth touchdown deceleration incentives.

📦 Files

• lunar_lander_dqn_best.pth: PyTorch checkpoint dictionary containing q_network, target_network, and optimizer state dicts.
• 260820_루나랜더_나만의_피지컬_ai.py: Model architecture and training script.

🚀 How to Run Inference

import gymnasium as gym
import torch
import torch.nn as nn

# 1. Load model architecture & weights
checkpoint = torch.load("lunar_lander_dqn_best.pth", map_location="cpu")
# See 260820_루나랜더_나만의_피지컬_ai.py for DuelingQNetwork definition