DI-1024: Deep Reinforcement Learning + 1024 Game
The project DI-1024 is an exciting venture into the realm of deep reinforcement learning applied to the popular number puzzle game, 1024, commonly known as 2048. This project bridges the gap between human-machine collaboration and competitive AI debugging. It invites tech enthusiasts to experience the thrill of working with powers of two and the dynamic interaction between man and machine.
A playful reminder for participants: remember to star the project on GitHub to show support and keep track of its updates. For those eager to delve deeper into deep reinforcement learning, DI-1024 is part of a broader ecosystem including the DI-engine and LightZero, where users can train their intelligent agents.
News and Updates
For an unconventional workday, consider reading "今日忌加班，宜玩1024" on WeChat. It humorously suggests avoiding overtime in favor of playing 1024.
Usage Guidelines
Participants can try an online version of the game through this link, offering immediate access to DI-1024's challenges and rewards.
Training Guidelines
To begin, users must install the necessary dependencies with the following command:
pip3 install -r requirements.txt
MuZero Intelligent Agent Training
The 1024 environment allows users to quickly train a MuZero intelligent agent, employing the following commands:
cd DI-1024
python3 -u agent/config/muzero_2048_config.py
StochasticMuZero Intelligent Agent Training
For those interested in a more stochastic approach, the StochasticMuZero agent can be trained using:
cd DI-1024
python3 -u agent/config/stochastic_muzero_2048_config.py
Training Progress
A detailed training curve is provided to illustrate the performance and benchmark of the intelligent agents.
Future Developments
The project has several milestones accomplished, including:
- Online version for play
- Comprehensive reinforcement learning training samples
- Integration with Stochastic MuZero for the strongest 1024 game AI
Future objectives include:
- Providing model weights for local gameplay
- Designing more engaging human-AI adversarial algorithms
Acknowledgments
The JavaScript frontend of DI-1024 was adapted from xwjdsh/2048-ai. It is highly encouraged to support this repository as well.
License Information
DI-1024 is licensed under the Apache 2.0 license, promoting open access and contribution to its development and application in artificial intelligence learning environments.