Decentralized multi-agent coordination across physical, financial, and topological networks presents a fundamental conflict between individual agent autonomy and global operational efficiency. Under heavy, non-Markovian traffic and long-range dependence, uncoordinated selfish routing induces severe herding behavior, yielding high Price of Anarchy (PoA) values, while centralized orchestration suffers from catastrophic curse-of-dimensionality bottlenecks. This paper introduces the Generalized Incentivized Mean Field (GIMF) architecture—a unified computational framework that synthesizes Mean Field Games (MFG), non-Markovian queueing and jump-diffusion dynamics, sub-Gaussian Multi-Armed Bandit (MAB) learning, and incentive-based Stackelberg control. Through an Adaptive Sequential Experimental Design and Statistical Meta-Synthesis (ASED-SMS) methodology, we execute a rigorous sequence of experiments across three distinct domain classes: Class A (Continuous Physical Mi-crogrids with Fractional Brownian Motion), Class B (Discontinuous DeFi Liquidity Pools with Heavy-Tailed Jump-Diffusions), and Class C (Scale-Free Topological Graph Networks with Non-Markovian Cascade Delays). Across all domain classes, the GIMF framework decouples computational 1 complexity from population size N , preserves an O(1/ √ N) mean-field approximation error, guarantees logarithmic exploration regret O(ln T), and consistently compresses the Price of Anarchy down to a Price of Stability (PoS) near unity (1.01 ≤ PoS ≤ 1.04).
- The Generalized Incentivized Mean-Field (GIMF) Framework: Adaptive Sequential Experimental Design and Statistical Meta-Synthesis Across Multi-Agent Domains
- Kemal Gursoy (Author) - Rutgers University, Management Science and Information Systems (RBS)
- INFORMS Annual Meeting (San Francisco, 11/01/2026–11/04/2026)
- Management Science and Information Systems (RBS)
- English
- Report
- 991032371566504646