distributed optimization

This study addresses the critical challenge of jointly optimizing system efficiency, individual comfort, and fairness in cost sharing within fully decentralized multi-agent collaboration—a key factor in preventing incentive misalignment and coordination failure. The work proposes the first decentralized coordination framework that simultaneously handles these three orthogonal objectives through a distributed optimization algorithm coupled with a preference-aware cost reallocation mechanism. Operating without centralized control and under bounded communication and computational overhead, the model achieves balanced multi-objective optimization. Empirical evaluation on two real-world datasets demonstrates the approach’s effectiveness: it maintains high system-wide performance while significantly improving individual satisfaction and fairness in cost distribution.

This study addresses the absence of a joint graph-structure design framework in existing decentralized optimization by proposing GATE. This method introduces a novel "hybrid-tearing" paradigm that decomposes the global problem into local subproblems via message passing, jointly designing constraint representations, dual variable blocks, and connected clusters to transcend conventional gossip communication limitations. Efficient collaborative computation is achieved through tree-recursive updates combined with a lightweight surrogate model (GATE-S). Theoretically, the algorithm is proven to attain a linear convergence rate under explicit dependencies on network topology and function regularity. Empirical evaluations further validate its significant advantages in reducing both communication and computational overhead.

本文通过线性规划方法，为任意信息网络和系统目标设计最优局部效用函数，以优化纯价格无政府状态（pPoA）性能保证。

To address the dual challenges of high communication overhead in large-scale federated learning and slow convergence of decentralized gossip protocols, this paper proposes a semi-decentralized probabilistic communication paradigm: agents dynamically select—according to an adjustable probability—either the central server or neighboring peers for communication, thereby jointly optimizing bandwidth efficiency and convergence speed. Methodologically, we integrate stochastic communication scheduling, multi-step local SGD, and gradient tracking into a unified framework for the first time, establishing a rigorous distributed optimization theory under non-convex and heterogeneous data settings; we prove linear-speedup convergence. Experiments demonstrate that our protocol significantly reduces the number of communication rounds while maintaining robustness across highly sparse topologies and strongly heterogeneous data distributions, offering a novel, efficient, and scalable paradigm for federated learning.

In multi-source heterogeneous systems (e.g., supply chains), primal, dual, and proximal agents coexist with immutable interfaces—yet existing consensus optimization methods, such as standard ADMM, assume agent homogeneity and cannot accommodate such heterogeneity. Method: We propose the first distributed consensus planning framework supporting collaborative optimization among all three agent types. By unifying linearized ADMM, dual ascent, and standard ADMM, we design a novel relaxation-and-enhancement update mechanism that accommodates structural mismatches without requiring interface modifications. Contribution/Results: Under mild assumptions, we establish rigorous convergence: O(1/k) rate under weak convexity and two-step linear convergence under strong convexity. Experiments on mixed-agent scenarios demonstrate both robustness and efficiency, validating the framework’s plug-and-play applicability. This work provides the first theoretically grounded, algorithmically practical solution for decentralized decision-making in heterogeneous multi-agent systems.

This work addresses the challenges of low coordination efficiency and limited connectivity in distributed online submodular maximization under communication delays. The authors propose the Distributed Online Greedy (DOG) algorithm, which integrates adversarial bandit learning with a delayed feedback mechanism to enable synchronized decision-making over arbitrary network topologies. For the first time, they establish a synchronous coordination framework in delayed distributed settings, revealing a fundamental trade-off between coordination performance and convergence time, and unifying the performance bounds of centralized and single-hop decentralized approaches. The DOG algorithm provides theoretical approximation guarantees characterized by network topology, achieving a controllable balance between coordination capability and response speed.

This work addresses the lack of tight and practical time complexity bounds in bandwidth-constrained decentralized non-convex stochastic optimization. By introducing graph-theoretic concepts—specifically the min-cut/max-flow quantity, Gomory-Hu trees, and Steiner tree packing—we develop a bandwidth-aware communication-computation co-analysis framework and derive a near-optimal lower bound on time complexity. Building on this foundation, we propose two algorithms: Grace SGD for homogeneous settings and Leon SGD for heterogeneous environments. Both algorithms achieve time complexity that is optimal up to logarithmic factors in their respective scenarios, significantly outperforming existing methods.

本文通过图模型设计解决在有限理性下人机协作问题，利用图子模型和变分法优化网络拓扑结构以最大化全局协调度。

This work addresses the challenge of efficiently computing Nash equilibria in multi-agent cooperative monitoring of dynamically relocatable targets under communication constraints, where the large combinatorial action space renders equilibrium computation intractable. The authors model target coverage as a defender-attacker zero-sum game and propose, for the first time, a distributed self-configuring communication-aware sensing framework that integrates communication constraints with submodular game theory. By combining a coordination-value mechanism with distributed bandit-submodular optimization, the approach efficiently approximates an approximate Nash equilibrium under stringent bandwidth limitations. Theoretical analysis and simulations demonstrate that the proposed method significantly outperforms baseline strategies, achieving coverage and game-theoretic payoffs markedly closer to optimal performance.

This work addresses the structural bias arising from strict block-wise decomposition when solving large-scale fixed-point equations in decentralized multi-agent systems. To mitigate this issue, the authors propose the Core-Halo decomposition method, which decouples write ownership from read context: each agent updates its core variables while reading overlapping halo variables. This design preserves parallelism while faithfully reconstructing the original problem. The approach innovatively introduces a Core-Halo structure to align operator dependencies, revealing the fundamental limitation of strict partitioning and establishing a lower bound on block-induced bias via a Bellman closure condition. Empirical results demonstrate that the framework achieves accuracy comparable to centralized solutions across diverse scenarios, while retaining the scalability and parallelism inherent to decentralized systems.