GNN-DT (Graph Neural Network Enhanced Decision Transformer) is a next-generation AI framework that seamlessly blends the power of Graph Neural Networks (GNNs) with Decision Transformers (DTs) to redefine decision-making in dynamic environments.

Read teh paper published in "Energy and AI":

• 📄 Full paper
• 📄 Preprint

🔹 Why GNN-DT?

• 🌟 Tackles scalability issues and sparse rewards
• 🔄 Adapts to ever-changing state-action spaces
• ⚡ Achieves unparalleled efficiency and robustness

💡 By leveraging the permutation-equivariant nature of GNNs and introducing an innovative residual connection mechanism, GNN-DT sets a new benchmark in optimization and AI-driven decision-making.

📌 This repository provides everything you need to explore and implement GNN-DT, including:

• ✅ Dataset generation
• ✅ Model training
• ✅ Evaluation scripts

Whether applied to electric vehicle (EV) charging optimization or other complex decision-making tasks, GNN-DT is your gateway to the AI-driven optimization. 🚀

image

• 📈 Learns from previously collected trajectories, reducing the need for extensive online interactions.
• 🎯 Effectively addresses the sparse rewards limitation of traditional RL algorithms.

• 🌍 GNN-based embeddings allow for effective adaptation to unseen environments.
• 🔄 Handles dynamic state-action spaces with varying numbers of entities over time.

• 🏆 Outperforms standard DT and RL baselines on real-world optimization tasks.
• 🚀 Requires significantly fewer training trajectories while achieving higher rewards.

• 🔢 Maintains performance across different problem sizes without retraining.
• ⚙️ Efficiently scales from small-scale to large-scale environments, as demonstrated in EV charging applications.

image

image

If you find this repository useful in your research, please cite our paper:

@misc{orfanoudakis2025gnndt,
 title={GNN-DT: Graph Neural Network Enhanced Decision Transformer for Efficient Optimization in Dynamic Environments}, 
 author={Stavros Orfanoudakis and Nanda Kishor Panda and Peter Palensky and Pedro P. Vergara},
 year={2025},
 eprint={2502.01778},
 archivePrefix={arXiv},
 primaryClass={cs.LG},
 url={https://arxiv.org/abs/2502.01778}, 
}

For any inquiries or contributions, feel free to open an issue or submit a pull request! 💡