Reproduce “Multi-Agent Reinforcement Learning for Power Grid Topology Optimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Erica van der Sar, Alessandro Zocca, Sandjai Bhulai
A hierarchical multi-agent reinforcement learning framework is presented for optimizing power grid topology through bus-switching actions, achieving robust and stable control. The approach, employing dependent learning agents with tuned hyperparameters, demonstrated competitive performance in the Grid2Op environment, outperforming independent multi-agent and single-agent methods in stability for specific algorithms.
Key insights
- Hyperparameter optimization is crucial for the performance of multi-agent reinforcement learning frameworks, especially for the SACD algorithm, as default single-agent parameters often yield suboptimal results in multi-agent settings.
- Dependent MARL strategies, specifically CTDE, offer superior learning stability compared to independent learning by allowing agents to implicitly or explicitly account for each other's actions during training.
- Consistent and rule-based policies for higher-level agents, such as ordering low-level agent activations, significantly improve the stability of learning for lower-level MARL agents compared to random activation sequences.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.