This mini-project delves into the application of Deep Q-learning, a powerful Deep Reinforcement Learning (DRL) technique, to train an AI agent to master the deceptively challenging Snake game. Our objective is to showcase the effectiveness of Deep Q-learning in enabling an agent to learn optimal strategies through interaction with the game environment and maximizing its reward. The project employs a custom-designed Deep Qlearning architecture within the agent, utilizing a neural network to approximate the Q-values associated with potential actions. This empowers the agent to dynamically evaluate and prioritize moves based on their predicted long-term value, navigating the intricate obstacles presented by the limited action space and fleeting rewards inherent to the game. By analyzing the agent's performance over numerous iterative interactions with the game, we witness its impressive evolution. The interplay of exploration and exploitation within the Deep Qlearning framework drives continuous improvement, culminating in the agent achieving exceptional performance marked by record-breaking scores. This remarkable journey not only demonstrates the power of Deep Q-learning in tackling complex optimization challenges but also paves the way for its broader application in diverse real-world domains where intelligent adaptation and decision-making are crucial.
This report presents a journey into the application of Deep Q-Learning (DQN) for mastering the classic Snake game. Our methodological exploration unfolds in three key chapters. First, we establish the theoretical foundations with an overview of Reinforcement Learning (RL) principles, including the Markov Decision Process framework and the Bellman Equation, which guides optimal policy discovery in Q-learning, the core algorithm at the heart of DQN. Next, we delve into the transformative power of DQN, delving into the intricacies of its core Q-learning mechanism and its enhanced ability to represent complex state and action spaces through deep neural networks. We then unravel the optimization process involved in training the DQN agent, encompassing loss functions and network updates that shape its decision-making capabilities. Finally, we bridge the gap between theory and practice with a detailed account of the experimental setup, encompassing the Snake game environment, the DQN agent's architecture, and the training procedures and evaluation metrics employed to assess its performance. The concluding chapter dissects the results, analyzing the effectiveness of DQN in mastering the Snake domain while acknowledging the remaining challenges, as the AI navigates increasingly complex scenarios. Ultimately, this work paves the way for further exploration in leveraging AI learning through games, unlocking new possibilities in the fascinating realm of RL.
The following libraries and frameworks are required for running the project:
- Python 3.10
- PyTorch
- OpenCV
- NumPy
- Matplotlib
- Scikit-learn
- h5py
- tqdm
- wandb ...
To install these dependencies, you can use the provided requirements.txt file.
pip install -r requirements.txt
-
Clone the repository:
git clone https://github.com/xt2201/it3320e-human-action-recognition-ucf101.git](https://github.com/sonho4ng/Snake-Game.git cd Snake-Game -
Install dependencies:
If you're using a
requirements.txtfile, run:pip install -r requirements.txtAlternatively, install required libraries individually using
pip.