Lux-Open

This is the official repo of the paper Emergent collective intelligence from massive-agent cooperation and competition.

Directory Struture

• agent : our agent implementation of centralized policy
• decentralized: our decentralized agent as part of the ablation study
• env: a wrapper of the Lux-AI Season 1 environment. See more in https://www.lux-ai.org/.
• lux: utility functions used to interact with the Lux backend. The original code is from https://github.com/Lux-AI-Challenge/Lux-Design-2021/tree/master/kits/python/simple.
• models: checkpoints we used to evaluation different stages of the training process.
• opponent: the NO.1 winner of the Lux-S1 kaggle challenge. The original code is at https://github.com/IsaiahPressman/Kaggle_Lux_AI_2021
• runEvaluation.py the main entrance for runnning experiements.

Usage

Environment Setup

• Install the Lux environment: see instructions at https://github.com/Lux-AI-Challenge/Lux-Design-2021
• pip install -r requirements.txt

Running experiements

• python runEvaluation.py
• Experiement Specification:
  • mode: evaluation mode, support self-play, different models and with opponent.
  • model1 and model2: evaluation model paths
  • num_games: num of games in the evaluation
  • map_size: evaluation mapsizes
  • seed: random seed of the env

Citation

@article{chen2023emergent,
  title={Emergent collective intelligence from massive-agent cooperation and competition},
  author={Chen, Hanmo and Tao, Stone and Chen, Jiaxin and Shen, Weihan and Li, Xihui and Cheng, Sikai and Zhu, Xiaolong and Li, Xiu},
  journal={arXiv preprint arXiv:2301.01609},
  year={2023}
}