Subject Ph. D.

Digital Twins (DTs) support complex distributed systems [1] such as wireless communication networks and Industry 4.0 environments. A DT maintains a virtual representation of the physical system, updated through observations and is widely used for monitoring and decision support. In wireless networks, Network DTs were recently formalized within standardization efforts [2], and are envisioned as key enablers for autonomous network control [3]. From an AI perspective, DTs comprise two complementary layers. The first relies on descriptive and predictive models to estimate the system state. Our previous work highlighted the difficulty of constructing such twins and maintaining a valid virtual-real mapping in industrial IoT [4]. The second layer concerns decision-making and control processes built on top of the DT, where action policies are derived from the virtual representation to influence the physical system behavior in a closed loop. In this setting, the DT becomes an active component of the perception-decision-action cycle. Multi-agent modeling and simulation capture these dynamics [5]. Autonomous control naturally aligns with agent-based abstractions, evolving from early theory [6] to Agentic AI [7]. In O-RAN and Edge settings, these frameworks enhance distributed resource allocation [8,9], typically leveraging Multi-Agent Reinforcement Learning (MARL) to optimize performance [10,11]. However, most existing agentic and learning-based approaches primarily emphasize performance maximization and implicitly assume a reliable representation of the environment. In practice, the internal models or representations used for decision-making, whether learned, simulated, or DT-based, may diverge from the physical system due to modeling assumptions, delayed observations, environmental changes, or non-stationary dynamics. Such divergence may not immediately affect performance, raising critical questions regarding the reliability and trustworthiness of agentic decision- making. As autonomy increases, agents must operate under partial observability, uncertainty, and evolving model fidelity [12]. When autonomous agents, such as those implemented using Federated Reinforcement Learning [13], act through an imperfect DT, they influence the system dynamics they observe. Decisions are therefore taken under incomplete or biased representations. A critical risk arises where good apparent performance coexists with incorrect internal reasoning, leading to "silent drift," instability, or unsafe behavior. Unlike approaches that optimize for "performance at all costs" [11], this thesis addresses the reliability and trustworthiness of the decision-making process itself. The central research question is : How can agentic AI systems make reliable and trustworthy decisions, without inducing unstable behaviors, when acting on dynamic wireless networks through an imperfect and evolving D T? Scientific Objectives This thesis aims to analyze agentic decision-making mechanisms in DT-based closed-loop control, using wireless networks as a representative class of complex dynamical systems. The main objectives are: -Trustworthy Control and Stability under Reality Gaps. Design agentic control mechanisms ensuring stable and safe behavior despite DT inaccuracies, by detecting uncertainty and enabling constrained decision-making or graceful degradation, building on safe reinforcement learning principles [14]. -Adaptive Causal World Modeling. Develop adaptive causal world models that distinguish structural environmental changes from modeling biases, ensuring valid decision logic under non-stationarity. -Scalable Decentralized Coordination. Analyze the limits of autonomy in large-scale networks and design decentralized and federated coordination mechanisms balancing local autonomy, global stability and frugality.

References [1] Sakhri et al. A digital twin-based energy-efficient wireless multimedia sensor network for waterbirds monitoring. Future Generation Computer Systems. 2024 Jun 1;155:146-63. [2] 3GPP TR 28.915, Study on Management Aspects of Network Digital Twin, Rel. 18, 2023. [3] Apostolakis et al. Digital twins for next-generation mobile networks: Applications and solutions. IEEE Communications Magazine. 2023 May 8;61(11):80-6. [4] Kherbache et al. Constructing a Network Digital Twin through formal modeling: Tackling the virtual-real mapping challenge in IIoT networks. Internet of Things. 2023 Dec 1;24:101000. [5] Shakya et al. MultiAgentNetSim: Empowering Next-Generation Network Modeling with Multi-Agent Simulation. IEEE Communications Magazine. 2024 Dec 23. [6] Wooldridge et al. Intelligent agents: Theory and practice. The knowledge engineering review. 1995 Jun;10(2):115-52. [7] Xi et al. The rise and potential of large language model based agents: A survey. Science China Information Sciences. 2025 Feb;68(2):121101. [8] Salama et al. Edge agentic ai framework for autonomous network optimisation in o-ran. In2025 IEEE 36th International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC) 2025 Sep 1 (pp. 1-7). IEEE. [9] O-RAN Alliance - Near-RT RAN Intelligent Controller (RIC) Architecture Specification. [10] Zhang et al. Multi-agent reinforcement learning in wireless distributed networks for 6g. arXiv preprint arXiv:2502.05812. 2025 Feb 9. [11] Toure et al. Multi-Objective Scheduling in Wireless Networks With Deep Reinforcement Learning. In2025 IEEE Wireless Communications and Networking Conference (WCNC) 2025 Mar 24 (pp. 1-6). IEEE. [12] Kaelbling et al. R. Planning and acting in partially observable stochastic domains. Artificial intelligence. 1998 May 1;101(1-2):99-134. [13] Ossongo et al. A multi-agent federated reinforcement learning-based optimization of quality of service in various LoRa network slices. Computer Communications. 2024 Jan 1;213:320-30. [14] Garc?a et al. A comprehensive survey on safe reinforcement learning. Journal of Machine Learning Research. 2015 Aug;16(1):1437-80.