NeuroBlock embeds the power of reinforcement learning by training an AI to play the classic Tetris game. Built using PyTorch for modeling and OpenCV for visualization, this project demonstrates how AI learns and adapts through gameplay.
- AI learns to play Tetris using reinforcement learning.
- Visual game interface powered by OpenCV to demonstrate the AI's performance.
- Customizable training parameters to fine-tune learning behavior.
numpytorchtermcoloropencv-pythonPillow
To install all dependencies, run:
pip install -r requirements.txt- Run the Pre-trained Model: View the AI in action with the final trained model:
python run.pyA window will open displaying the AI playing Tetris.
- Train the Model: Start training the model from scratch:
python train.pyYou can customize the following parameters in the code to optimize the training process:
| Parameter | Default Value | Description |
|---|---|---|
epochs |
30000 |
Number of training epochs. |
epsilon |
1 |
Initial epsilon value for the Epsilon-Greedy Algorithm. Should remain between 1 and the epsilon floor (0.001). |
gamma |
0.999 |
Epsilon decay rate. Higher values slow down the decay, while smaller values decrease epsilon more quickly. |
replay_size |
100000 |
Capacity of the replay buffer used to store Tetris states for training. |
minibatch_size |
200 |
Size of the minibatch sampled from the replay buffer during training. |
During gameplay, the AI's decisions and movements are displayed in real-time. Below is a sample frame of the Tetris game being played by the AI:
- Reinforcement Learning: The AI leverages Q-learning with a neural network to maximize its score by placing Tetris efficiently.
- Epsilon-Greedy Algorithm: Balances exploration and exploitation during gameplay to improve decision-making.
- Replay Buffer: Stores past game states to train the AI efficiently by replaying critical scenarios.
- Explore additional reinforcement learning strategies for better performance.
- Add a user interface to adjust training parameters dynamically.
Feel free to contribute or provide feedback! 🚀