https://github.com/starlitdreams/lunar-landing
This project implements a DQN agent using PyTorch to solve the LunarLander-v2 environment from OpenAI Gym. The agent learns to control the lunar lander using experience replay and a target network, aiming to maximize rewards by landing smoothly. Uses CUDA for computation.
https://github.com/starlitdreams/lunar-landing
artificial-intelligence cuda deep-learning gymnasium neural-network neural-networks numpy nvidia-gpu python python3 torch
Last synced: 7 months ago
JSON representation
This project implements a DQN agent using PyTorch to solve the LunarLander-v2 environment from OpenAI Gym. The agent learns to control the lunar lander using experience replay and a target network, aiming to maximize rewards by landing smoothly. Uses CUDA for computation.
- Host: GitHub
- URL: https://github.com/starlitdreams/lunar-landing
- Owner: StarlitDreams
- License: mit
- Created: 2024-09-04T11:43:08.000Z (about 1 year ago)
- Default Branch: main
- Last Pushed: 2024-09-17T13:22:39.000Z (about 1 year ago)
- Last Synced: 2025-03-30T03:31:45.298Z (7 months ago)
- Topics: artificial-intelligence, cuda, deep-learning, gymnasium, neural-network, neural-networks, numpy, nvidia-gpu, python, python3, torch
- Language: Python
- Homepage:
- Size: 250 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Lunar Lander - Deep Q-Network (DQN)
This repository contains the implementation of a Deep Q-Network (DQN) agent that solves the **LunarLander-v2** environment from OpenAI Gym using PyTorch.
## Table of Contents
- [Introduction](#introduction)
- [Project Structure](#project-structure)
- [Dependencies](#dependencies)
- [Installation](#installation)
- [Training](#training)
- [Results](#results)
- [Contribution](#contribution)
- [License](#license)
## Introduction
This project implements a DQN agent to solve the **LunarLander-v2** environment using PyTorch and OpenAI Gym. The agent uses experience replay and a target network to improve the stability of the learning process.
The goal is for the agent to land the lunar lander on the target surface as smoothly and efficiently as possible, maximizing the cumulative rewards over episodes.
## Project Structure
```
├── main.py # Contains the code
├── README.md # This readme file
├── checkpoint.pth # Model weights after training (saved checkpoint)
└── trained_agent_video.mp4 # Video recording of the trained agent
```
## Dependencies
The project requires Python 3.x and the following libraries:
- `torch`
- `numpy`
- `gymnasium`
- `IPython`
- `imageio`
- `random`
You can install all dependencies using the following command:
```bash
pip install numpy gymnasium ipython imageio
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
pip install --upgrade cryptography
pip install --upgrade wheel setuptools
pip install gymnasium[box2d]
```
## Installation
1. Clone the repository:
```bash
git clone https://github.com/your_username/lunar-lander-dqn.git
cd lunar-lander-dqn
```
2. Install the dependencies:
```bash
pip install -r requirements.txt
```
Alternatively, if you don't have a `requirements.txt` file, you can install the necessary dependencies as shown in the Dependencies section.
3. Run the code:
```bash
python main.py
```
## Training
To train the agent from scratch, run:
```bash
python main.py
```
The DQN agent will interact with the environment over multiple episodes, updating its policy based on the experience replay and target networks. The training process will output the progress every 100 episodes and save a checkpoint of the model when the environment is solved (when the average score reaches 200.0 over 100 episodes).
You can adjust the hyperparameters (e.g., learning rate, batch size) by modifying the constants defined in `main.py`.
## Results
During the training process, the agent’s performance will be tracked. Once the average score over 100 episodes reaches 200.0, the environment is considered solved, and the model weights will be saved to `checkpoint.pth`.
Sample output:
## Watch the Agent Play
A pre-trained model has been provided. To see the agent play the game, you can watch the video by running the code or viewing the `video.mp4` file.
You can generate a new video of the trained agent using the following function:
```python
show_video_of_model(agent, 'LunarLander-v2')
```
The video will be saved as `trained_agent_video.mp4` in the project directory.
## Contribution
Contributions are welcome! Please feel free to submit a Pull Request or open an Issue for any improvements or bug fixes.
## License
This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details.