Abstract:
Multi-agent systems (MAS) are increasingly adopted in various domains, necessitating efficient decision-making strategies that balance individual agent preferences with overarching system objectives. This paper presents a novel framework for Multi-Agent Multi-Criteria Decision-Making (MAS-MCDM), which extends any MCDM algorithm to address the challenges of scalability, resource constraints, and inter-agent interactions. We propose four distinct decision-making approaches: 1. Full Enumeration, which guarantees the global optimum but is computationally prohibitive for large-scale problems. 2. Independent Agent Decomposition, which enhances scalability but neglects inter-agent synergies. 3. Iterative Refinement, where agents adjust decisions dynamically based on system-level feedback to satisfy constraints. 4. Iterative Refinement with Inter-Agent Interactions, which integrates cooperation and competition dynamics through an interaction matrix. To validate the framework, we adapt the Rastrigin function — traditionally used for single-agent global optimization — into a discrete MAS-MCDM benchmark. By transforming the continuous optimization problem into a decision-making task, we examine how constraints and interactions influence solutions, shifting them away from the global minimum at (x=0). The experiments demonstrate that iterative refinement approaches effectively navigate resource limitations, while interaction-aware models enable the emergence of cooperative and competitive behaviors. The results highlight trade-offs between optimality, scalability, and inter-agent coordination, providing insights into designing robust multi-agent systems for multi-criteria decision-making (MAS-MCDM) strategies in real-world applications. Future directions include refining interaction models, integrating reinforcement learning, and extending applications to autonomous and resource-constrained systems.
Keywords:Multi-Agent Systems (MAS), Multi-Criteria Decision-Making (MCDM), optimization framework, inter-agent interactions, benchmarking with Rastrigin function.
Fulltext:
PDF file (547 kB)