Lunar Lander Agent using Deep Q-Network

• Developed a Deep Q-Network (DQN) agent from scratch to solve the LunarLander-v3 environment using PyTorch and Gymnasium, demonstrating a strong grasp of reinforcement learning concepts.
• Designed and implemented a custom neural network architecture to accurately approximate Q-values, enabling the agent to make optimal decisions in a complex, continuous state space.
• Built an efficient experience replay memory system to store and sample past experiences, significantly improving training stability and sample efficiency.
• Engineered a robust DQN agent class, integrating neural network optimization, soft target network updates, and epsilon-greedy exploration for effective learning.
• Tuned hyperparameters and conducted extensive training over thousands of episodes, achieving high average scores and solving the environment ahead of the benchmark.
• Visualized the agent’s learning progress and final performance by generating and displaying gameplay videos, providing clear evidence of the model’s effectiveness.