Reproduce “Fault Tolerant Multi-Agent Learning with Adversarial Budget Constraints” 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

David Mguni, Yaqi Sun, Haojun Chen, Wanrong Yang, Amir Darabi, Larry Olanrewaju Orimoloye, Yaodong Yang

Researchers from Queen Mary University London and Peking University introduce MARTA (Multi-Agent Robust Training Algorithm), a framework that trains cooperative multi-agent policies to tolerate agent malfunctions by selectively inducing faults during training. This method consistently enhances performance and reduces failure rates across various tasks, integrating as a plug-and-play layer without modifying base MARL algorithms.

Key insights

  • The core of the system lies in its selective adversarial fault induction, where the Switcher's cost-aware or budget-constrained decisions guide training pressure to critical situations, balancing robustness with nominal performance.
  • MARTA's design is "plug-and-play," enabling seamless integration with existing MARL backbones like QMIX and VDN without requiring architectural modifications.
  • The theoretical analysis demonstrates that the fault-switching Markov game admits a unique minimax value and that the associated Bellman operator is a contraction, ensuring the convergence of a Q-learning-type scheme.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.