Multi-Agent Robust Training (MARTA)

• MARTA is a robust multi-agent training method that employs auxiliary perturbations to simulate adversarial noise and ensure decentralized policies remain effective.
• It systematically introduces variations in communication, observations, transitions, and teammate actions to test system resilience under non-nominal conditions.
• This approach enhances policy stability and provides a baseline for developing resilient multi-agent reinforcement learning systems in complex environments.

Searching arXiv for "Multi-Agent Robust Training Algorithm MARTA" and closely related robust multi-agent training papers. Tool call: arxiv_search(query="\"Multi-Agent Robust Training Algorithm\" OR MARTA robust multi-agent training", max_results=10, sort_by="relevance") Searching for papers on "robust multi-agent reinforcement learning" and "MARTA" as an acronym or named method. Tool call: arxiv_search(query="robust multi-agent reinforcement learning MARTA", max_results=10, sort_by="relevance") Multi-Agent Robust Training Algorithm (MARTA) is best understood, in the robust multi-agent learning literature, as a MARTA-style robustification baseline in which an ego system is trained under auxiliary perturbations or adversarial noise so that decentralized policies remain effective when communication, observations, transitions, or teammate actions deviate from nominal training