Publications

A hierarchical agentic incident-response framework that combines LLM-based attack inference, rollout planning, and digital-twin validation to improve reliable recovery from multi-stage cyberattacks.

An LLM-guided constrained reinforcement learning method for scalable cloud-edge task offloading and resource allocation, combining action demonstrations, reward shaping, and probabilistic action sampling.

CORAL formulates in-context reinforcement learning as emergent communication, using a pre-trained world-model agent to provide transferable context for sample-efficient and zero-shot adaptation.

An end-to-end LLM agent for autonomous network incident response that unifies perception, attack-model reasoning, response planning, and action with in-context adaptation.

A delay-aware deep RL framework for O-RAN control that uses LSTM-based delay prediction and temporal reward realignment to stabilize learning under stochastic feedback latency.

A distributed online convex optimization framework for networked decision-making with nonseparable costs and constraints, enabling coordinated learning under coupled objectives.

A Stackelberg-game framework for tackling the deadly triad in off-policy deep reinforcement learning, which models representation learning and Q-learning as two strategic agents in a hierarchical game

A Hessian-free meta-learning algorithm for ergodic Linear Quadratic Regulator (LQR) tasks is introduced, which directly approximates the meta-policy gradient with zeroth-order information via Gaussian smoothing and Stein’s identity.

A comprehensive discussion on the role of game-theoretic models and foundation models in cyber deception operations within strategic cyber warfare, highlighting their symbiotic relationship in automated cyber defense mechanisms.

An urban traffic digital twin system that includes AI-powered traffic sensing and event-triggered adaptive calibration, which provides driver-centric situational awareness and predicts network-wide mobility and safety risks in real time.

This paper establishes a Bayesian persuaded branching process to study social network platforms' (SNP) tagging policy design under misdetection of misinformation.

A novel Conjectural Online Learning (COL) method for adaptive security response strategies in IT infrastructures, addressing model misspecification and uncertainty through Bayesian learning and rollout.

A physics-informed reinforcement learning (PIRL) approach leveraging digital twins for efficient and generalizable wireless indoor navigation using mmWave signals, achieving zero-shot generalization in unseen environments.

An AI-powered pan-tilt camera-based traffic sensing system, composed of spatial-temporal graph transformer for real-time traffic estimation and camera control through distributed online convex optimization.

A proactive end-to-end security mechanism for 5G-enabled multi-domain communication networks, utilizing decision-dominant learning to prevent lateral movement of adversaries within a zero-trust architecture.

A game-theoretic framework for distributed empirical risk minimization (ERM) problems over networks where nodes strategically choose both learning parameters and network structure to optimize performance.

A comprehensive examination of multi-agent learning (MAL) from the perspective of information structures, proposing a taxonomy based on information, belief, and policy generation mechanisms.

A comprehensive review of game-theoretic multiagent decision making over networks, developing a unified dynamical-system viewpoint of reinforcement learning and control & optimization over networks, highlighting its applications and the growing interest in complex network systems.

A directional compactly supported box spline tight framelet with simple geometric structure is built such that all the high-pass filters in its underlying tight framelet filter bank have only two nonzero coefficients with opposite signs.