Selected work

Representative Papers

Peer Effect Estimation in the Presence of Simultaneous Feedback and Unobserved Confounders

Aug 05, 2025

This paper addresses the dual challenges of simultaneity bias and unobserved confounding in estimating peer causal effects in complex systems such as social networks. Methodologically, it proposes the first unified framework jointly tackling both issues: (i) an I-G matrix transformation to decouple mutual dependencies; (ii) network-derived instrumental variables constructed via a two-stage residual inclusion (2SRI) approach; and (iii) an adversarial debiasing mechanism to correct for double bias under nonlinear, high-dimensional settings. Theoretically, the estimator is proven consistent under standard regularity conditions. Empirically, it significantly outperforms existing methods on semi-synthetic and real-world datasets, yielding more accurate recovery of true peer effects. The core contribution lies in the first formal unification of feedback and confounding modeling, establishing a deep learning–instrumental variable hybrid paradigm for network causal inference that offers both theoretical guarantees and strong empirical performance.

0 citationsRead paper

Multi-Agent Reinforcement Learning for Resources Allocation Optimization: A Survey

Apr 29, 2025

This paper addresses the challenges of Resource Allocation Optimization (RAO) in dynamic, decentralized environments. To tackle these challenges, it systematically surveys state-of-the-art applications of Multi-Agent Reinforcement Learning (MARL) to RAO. We propose the first three-dimensional taxonomy for RAO—spanning collaboration structure, communication mechanism, and learning paradigm—unifying over 120 recent works and constructing a comprehensive technical landscape across key domains including network slicing, edge computing, and smart grids. By integrating mainstream MARL methodologies—including value decomposition, policy gradient methods, communication-aware learning, and opponent modeling—we establish a method-to-use-case mapping framework. Furthermore, we release an open, continuously updated MARL-RAO research roadmap, accompanied by a technology selection guide and a practical evaluation framework. Our work significantly enhances the deployability of RAO solutions in real-world systems, improving scalability, robustness, and operational feasibility.

0 citationsRead paper

Comparative Analysis of POX and RYU SDN Controllers in Scalable Networks

Mar 28, 2025International journal of Computer Networks & Communications

This study systematically evaluates the Quality-of-Service (QoS) performance differences between two prominent open-source SDN controllers—POX and Ryu—in scalable network environments. Using Mininet, we construct multi-scale topologies and implement OpenFlow-based flow programming and Python-based controller logic to quantitatively measure key QoS metrics: throughput, end-to-end latency, and jitter. Our work presents the first cross-topology empirical quantification of their scalability boundaries. Results show that Ryu achieves 42% higher throughput and 31% lower average latency than POX at the thousand-node scale, demonstrating superior production-readiness for large deployments. Conversely, POX exhibits advantages in small-scale scenarios—including faster startup time and greater debugging flexibility—due to its lightweight architecture. These findings provide data-driven, practical guidance for SDN controller selection and optimization in real-world network deployments.

0 citationsRead paper

Nucleolus Credit Assignment for Effective Coalitions in Multi-agent Reinforcement Learning

Mar 01, 2025

In cooperative multi-agent reinforcement learning (MARL), monolithic global coalition formation leads to inaccurate credit assignment, poor task decomposition capability, and suboptimal performance. To address this, this paper introduces the *nucleolus*—a solution concept from cooperative game theory—into MARL credit assignment for the first time, proposing the Nucleolus Q-Learning framework. Our method automatically identifies stable and efficient small-scale subcoalitions via the nucleolus solution, enabling interpretable subtask decomposition while providing theoretical guarantees on convergence and stability. Evaluated on Predator-Prey and StarCraft II benchmarks across multiple difficulty levels, our approach achieves significant improvements in win rate and cumulative reward—particularly outperforming four state-of-the-art baselines in hard and super-hard scenarios. These results empirically validate the effectiveness and generalizability of multi-subcoalition structures for modeling complex cooperative tasks.

0 citationsRead paper

Recent publications

Latest Papers

Peer Effect Estimation in the Presence of Simultaneous Feedback and Unobserved Confounders

Aug 05, 2025

This paper addresses the dual challenges of simultaneity bias and unobserved confounding in estimating peer causal effects in complex systems such as social networks. Methodologically, it proposes the first unified framework jointly tackling both issues: (i) an I-G matrix transformation to decouple mutual dependencies; (ii) network-derived instrumental variables constructed via a two-stage residual inclusion (2SRI) approach; and (iii) an adversarial debiasing mechanism to correct for double bias under nonlinear, high-dimensional settings. Theoretically, the estimator is proven consistent under standard regularity conditions. Empirically, it significantly outperforms existing methods on semi-synthetic and real-world datasets, yielding more accurate recovery of true peer effects. The core contribution lies in the first formal unification of feedback and confounding modeling, establishing a deep learning–instrumental variable hybrid paradigm for network causal inference that offers both theoretical guarantees and strong empirical performance.

0 citationsRead paper

Multi-Agent Reinforcement Learning for Resources Allocation Optimization: A Survey

Apr 29, 2025

This paper addresses the challenges of Resource Allocation Optimization (RAO) in dynamic, decentralized environments. To tackle these challenges, it systematically surveys state-of-the-art applications of Multi-Agent Reinforcement Learning (MARL) to RAO. We propose the first three-dimensional taxonomy for RAO—spanning collaboration structure, communication mechanism, and learning paradigm—unifying over 120 recent works and constructing a comprehensive technical landscape across key domains including network slicing, edge computing, and smart grids. By integrating mainstream MARL methodologies—including value decomposition, policy gradient methods, communication-aware learning, and opponent modeling—we establish a method-to-use-case mapping framework. Furthermore, we release an open, continuously updated MARL-RAO research roadmap, accompanied by a technology selection guide and a practical evaluation framework. Our work significantly enhances the deployability of RAO solutions in real-world systems, improving scalability, robustness, and operational feasibility.

0 citationsRead paper

Comparative Analysis of POX and RYU SDN Controllers in Scalable Networks

Mar 28, 2025International journal of Computer Networks & Communications

This study systematically evaluates the Quality-of-Service (QoS) performance differences between two prominent open-source SDN controllers—POX and Ryu—in scalable network environments. Using Mininet, we construct multi-scale topologies and implement OpenFlow-based flow programming and Python-based controller logic to quantitatively measure key QoS metrics: throughput, end-to-end latency, and jitter. Our work presents the first cross-topology empirical quantification of their scalability boundaries. Results show that Ryu achieves 42% higher throughput and 31% lower average latency than POX at the thousand-node scale, demonstrating superior production-readiness for large deployments. Conversely, POX exhibits advantages in small-scale scenarios—including faster startup time and greater debugging flexibility—due to its lightweight architecture. These findings provide data-driven, practical guidance for SDN controller selection and optimization in real-world network deployments.

0 citationsRead paper

Nucleolus Credit Assignment for Effective Coalitions in Multi-agent Reinforcement Learning

Mar 01, 2025

In cooperative multi-agent reinforcement learning (MARL), monolithic global coalition formation leads to inaccurate credit assignment, poor task decomposition capability, and suboptimal performance. To address this, this paper introduces the *nucleolus*—a solution concept from cooperative game theory—into MARL credit assignment for the first time, proposing the Nucleolus Q-Learning framework. Our method automatically identifies stable and efficient small-scale subcoalitions via the nucleolus solution, enabling interpretable subtask decomposition while providing theoretical guarantees on convergence and stability. Evaluated on Predator-Prey and StarCraft II benchmarks across multiple difficulty levels, our approach achieves significant improvements in win rate and cumulative reward—particularly outperforming four state-of-the-art baselines in hard and super-hard scenarios. These results empirically validate the effectiveness and generalizability of multi-subcoalition structures for modeling complex cooperative tasks.

0 citationsRead paper