Categorizes multi-agent system decision-making into rule-based, game theory, evolutionary, MARL, and LLM-based types, emphasizing MARL and LLM advantages over traditional methods.
A Comprehensive Survey on Multi-Agent Cooperative Decision-Making: Scenarios, Approaches, Challenges and Perspectives categorizes multi-agent system (MAS) decision-making approaches into five broad types: rule-based (primarily fuzzy logic), game theory-based, evolutionary algorithms-based, deep multi-agent reinforcement learning (MARL)-based, and large language model (LLM) reasoning-based methods . The survey emphasizes that while traditional methods like rule-based, game theory, and evolutionary algorithms show some effectiveness, they often rely heavily on pre-defined strategies and assumptions, limiting their adaptability in dynamic and uncertain environments . In contrast, MARL-based and LLM-based decision-making methods demonstrate significant advantages due to their ability to learn, adapt, and handle complex, real-time interactions . As a result, the paper focuses extensively on these two paradigms, providing in-depth analysis of their methodologies, taxonomies, benefits, and limitations . Applications of these advanced methods span autonomous driving, drone navigation, disaster rescue, and collaborative robotics
This survey provides a systematic taxonomy of Multi-Agent Cooperative Decision-Making (MACDM), categorizing the landscape into five distinct paradigms: rule-based systems, game-theoretic approaches, evolutionary algorithms, Multi-Agent Reinforcement Learning (MARL), and emerging Large Language Model (LLM)-based agents. The work offers a comparative analysis of these methodologies, evaluating their performance across diverse application scenarios ranging from autonomous driving to robotic swarms. By dissecting the architectural foundations of each approach, the authors illustrate the historical progression from rigid, heuristic-driven models to more dynamic, learning-based frameworks capable of handling complex, stochastic environments.
A key contribution of this material is its emphasis on the transformative potential of MARL and LLM-based approaches over traditional methods. It details how MARL addresses the scalability and adaptability limitations of game theory and rule-based systems by enabling agents to learn optimal policies through interaction. Furthermore, the survey explores the disruptive role of LLMs in MAS, highlighting their capacity for semantic reasoning, natural language communication, and zero-shot generalization, which facilitate more human-like cooperation and reduce the reliance on extensive reward engineering.
This research is vital for understanding the current frontier of distributed intelligence, as it identifies critical challenges such as credit assignment, communication bottlenecks, and safety alignment in cooperative settings. By outlining future perspectives, the paper serves as a roadmap for researchers aiming to bridge the gap between symbolic reasoning and sub-symbolic learning. It underscores a significant industry shift toward integrating generative AI with reinforcement learning, paving the way for more robust and interpretable multi-agent systems.
# Summary of "A Comprehensive Survey on Multi-Agent Cooperative Decision-Making"
This survey provides a structured taxonomy of multi-agent cooperative decision-making (MCDM) systems, categorizing approaches into rule-based, game theory-based, evolutionary methods, Multi-Agent Reinforcement Learning (MARL), and Large Language Model (LLM)-based frameworks. It highlights the evolution from traditional methods—such as rule-based systems and game-theoretic solutions—to more adaptive, learning-driven paradigms like MARL and LLM-enhanced decision-making. The paper emphasizes that while classical approaches offer interpretability and theoretical guarantees, modern techniques leverage scalability, generalization, and contextual reasoning, particularly in dynamic and uncertain environments.
The key contributions include: 1. Taxonomy and Comparison: A clear framework for classifying MCDM methods, showcasing their strengths and limitations across different scenarios (e.g., robotics, IoT, financial markets). 2. MARL and LLM Advantages: MARL excels in handling complex, decentralized interactions with delayed rewards, while LLMs introduce natural language understanding and few-shot adaptability, mitigating the need for extensive retraining. 3. Open Challenges: The survey identifies critical gaps, such as scalability, generalization across domains, robustness to adversarial behavior, and ethical alignment in cooperative AI systems.
This work matters because it synthesizes decades of research while pointing toward future directions, particularly the integration of LLMs with MARL to bridge the gap between symbolic reasoning and data-driven decision-making. For researchers and practitioners, it serves as a roadmap for selecting or advancing MCDM techniques in real-world applications. For the broader AI community, it underscores the shift toward collaborative, intelligent systems that can operate in open-ended, human-aligned environments.
Source: [arXiv:2503.13415v1](https://arxiv.org/html/2503.13415v1)