Snake Game using Reinforcement Learning

Summary:

In this project, we explored reinforcement learning (RL) techniques by developing agents to play the classic Snake game. We implemented three different agents: DQN (with target network), Double DQN, and Curriculum Learning. The project demonstrated the effect of hyperparameter tuning and various network architectures. The agents were trained using a state representation vector of 11 features to guide their movements toward the food while avoiding collisions

Tools & Technologies:

• Programming Language and Tools: Python, PyTorch, Pygame
• Reinforcement Learning: DQN, Double DQN, Curriculum Learning

Key Contributions:

• Developed and evaluated RL models, including DQN with experience replay, Double DQN, and Curriculum Learning, in both 10x10 and 20x20 grids
• Integrated a distance-based reward function to optimize agent behavior, enhancing policy learning and performance
• Implemented batch training and compared performance across various configurations, achieving up to a maximum score of 61 in the 20x20 grid using DQN with a freeze target network

Results:

Curriculum learning yielded the best performance on smaller grids, while DQN with a freeze target network excelled on larger grids. These results provided insights into how RL techniques handle complex, dynamic environments like Snake