Skip to content

Releases: jaymarichua/SuperStar

SuperStar: Superhuman Reinforcement Learning Agent for StarCraft II

30 Dec 06:40
@jaymarichua jaymarichua
v1.0.0
c4284f1
Compare
Choose a tag to compare
SuperStar: Superhuman Reinforcement Learning Agent for StarCraft II Latest
Latest

SuperStar: Configurable Real-Time AI for StarCraft II with APM Control and Future Reward-Hacking Extensions

Description

SuperStar is a configurable, real-time AI built for StarCraft II, inspired by DI-Star’s large-scale reinforcement learning pipelines. This new release introduces a framework for throttling an AI agent’s actions per minute (APM), ensuring more balanced matches against human players. It also lays the groundwork for advanced “reward-hacking” experimentation to explore novel AI behaviors in the near future.

SuperStar leverages training methodologies proven to reach grandmaster-level performance in Zerg vs. Zerg, while remaining highly adaptable for researchers or enthusiasts seeking to test custom RL strategies, supervised learning expansions, and soon, novel reward mechanisms.

Changelog

• Configurable Actions Per Minute (APM) Setting (2024-12-29)

  • Allows users to limit the bot’s effective speed.
  • Prevents overwhelming micro in human-vs-AI matches.

• Reward-Hacking Hooks (Coming Soon)

  • Future features planned to manipulate or visualize unusual reward structures.
  • Encourages deeper experimentation with AI behaviors.

Legacy DI-Star Features

• Pre-Trained Zerg vs. Zerg Agents

  • Immediate out-of-the-box gameplay demos.
  • Includes scripts for quick tests or human challenge matches.

• SL & RL Training Framework

  • Scalable codebase for large or single-PC setups.
  • Updated super/master training code references from 2022.

• Harstem Exhibitions & YouTube Matches

  • Demonstrations of the AI’s high-level gameplay against a professional StarCraft II caster and player.

• Ongoing Improvements

  • Roadmap includes stronger RL baselines, expanding from single-race matchups to multi-race integration.

Getting Started

  1. Installation

    • Clone this repository and follow the dependency requirements as listed in the README.
    • Ensure StarCraft II or the PySC2 environment is correctly set up.

  2. Configuration

    • Edit actor.bot_target_apm in the config to define a maximum number of actions per minute for the AI.
    • Adjust existing environment, race, or training parameters based on your hardware or research goals.

  3. Demo & Play

    • Use bin/play.py or other provided scripts to watch AI matches or play against the AI directly.
    • Explore different APM caps to observe how the AI’s micro and macro performance scale with speed constraints.

  4. Future Reward-Hacking

    • Keep an eye on our documentation for future expansions that let you alter the reward structure or run advanced experiments in emergent AI behaviors.

Tags & Assets

  • Tag: v1.0.0-superstar
  • Assets: Pre-trained Zerg agent, environment configs, training scripts, documentation.
  • Next Steps:
    • Fine-tune RL or SL agents with the newly configurable APM.
    • Experiment with upcoming reward-hacking hooks for unconventional AI strategies.

Credits & Contact

SuperStar is built upon foundational elements from DI-Star, a well-established platform for SC2 AI research. We welcome feedback, pull requests, and discussions about further collaborations or new features. Please open issues or send pull requests in this repository to get involved.

Assets 2
Loading