DI-star: A Cutting-edge Game AI Training Platform
Introduction
DI-star is an innovative platform specifically designed for training artificial intelligence in the realm of StarCraft II, a challenging and popular real-time strategy game. The platform aims to enable large-scale distributed training, and it has already achieved the impressive feat of creating grandmaster-level AI agents. This project provides a variety of resources and tools to support AI development and experimentation in gaming scenarios.
Key Features
-
Play Demo and Test Code: DI-star offers users the opportunity to test and interact with pre-developed AI agents, which allows them to directly see the capabilities and performance of the AI in StarCraft II environments.
-
Pre-Trained Agents: The platform includes initial versions of Zerg versus Zerg pre-trained agents, utilizing both supervised learning (SL) and reinforcement learning (RL) methodologies. These agents offer a solid foundation for further training and development.
-
Training Codes and Guides: DI-star provides comprehensive training codes for both SL and RL, including a detailed guide for training with minimal resources, such as a single PC. This makes training accessible to those with limited computational resources.
-
Enhanced AI Battles: The trained agents have been tested in battles with human players, such as the well-known player Harstem, and plans are underway to develop even more robust RL agents.
Getting Started
To start using DI-star, you need to set up your system as follows:
-
Install StarCraft II: Download and install the retail version of StarCraft II. Ensure that the installation path is set up correctly in your system’s environment variables.
-
Install DI-star and PyTorch: Clone the DI-star repository from GitHub and install it locally. Also, install PyTorch (version 1.7.1 with CUDA is recommended) to facilitate the AI training processes.
-
Hardware Requirement: A GPU is recommended for optimal performance during real-time agent tests.
Interaction with Pre-Trained Agents
-
Download the Correct StarCraft II Version: Ensure you have version 4.10.0, as this version is optimized for performance with the trained AI models.
-
Model Downloads: Fetch the appropriate AI models depending on your needs; options include models trained with human replays or reinforcement learning. These models can simulate strategies from different player skill levels, such as diamond players and grandmasters.
-
Agent Testing: Users can engage in several testing scenarios, such as playing against the AI agent, observing AI versus AI matches, or having the AI compete against built-in game bots.
Building and Training Custom Agents
DI-star is not just a platform for playing against pre-trained agents; it also supports the development of custom AI agents. Users can build different agents within the existing framework, allowing for experimentation and innovation in AI strategy development. Whether through supervised learning or reinforcement learning, the framework provides robust training support and distributes training capabilities for enhanced scalability.
Community and Support
DI-star encourages community interaction and provides support through platforms like Slack and Discord, where users can share insights, seek help, and collaborate on AI development projects.
Conclusion
DI-star is a powerful platform for researchers, developers, and enthusiasts interested in AI and gaming. It provides all necessary tools to explore the intricate world of StarCraft II AI, from pre-trained agents to custom agent development, supported by community interactions and comprehensive documentation. With DI-star, advancing the AI capabilities in StarCraft II has never been more accessible or exciting.