2048 DeepQ Agent

• Tech Stack: Python, PyTorch, Reinforcement Learning
• Github URL: Project Link

Developed a Deep Q-Learning agent using PyTorch that autonomously learns and optimizes strategies to achieve high scores in a 2048 game environment

Achievments

• Control Phase: Achieved a control score of approximately 6,500 with the agent reaching the 2048 tile once every 1,000 games
• Optimized Phase: Enhanced performance to reach scores up to 50,000, with the agent obtaining the 2048 tile in 70% of the games.

Features

• Deep Q-Network (DQN): Utilizes a convolutional neural network to approximate Q-values for state-action pairs
• Experience Replay: Implements a replay memory to store and sample past experiences, improving learning stability
• Epsilon-Greedy Strategy: Balances exploration and exploitation during training, with dynamic epsilon decay
• Target Network: Incorporates a separate target network to stabilize training by reducing correlations between target and prediction
• One-Hot Encoding: Converts the game board into a tensor representation suitable for neural network input