Faculty of Science, Technology and Medicine
Faculty of Law, Economics and Finance
Faculty of Humanities, Education and Social Sciences
Interdisciplinary Centre for Security, Reliability and Trust
Luxembourg Centre for Systems Biomedicine
Luxembourg Centre for Contemporary and Digital History
Luxembourg Centre for European Law
Luxembourg Centre for Socio-Environmental Systems
The Doctoral School in Science and Engineering is happy to invite you to HE Ke’s defence entitled
RISK-AWARE INTELLIGENCE FOR LARGE-SCALE PARTIALLY OBSERVABLE COMMUNICATION SYSTEMS
Supervisor: Prof Symeon CHATZINOTAS
The transition to Sixth Generation (6G) communication systems introduces unprecedented scale, decentralization, and dynamism, exemplified by massive MIMO terrestrial networks and ultra-dense Low Earth Orbit (LEO) satellite constellations. In such environments, acquiring complete and timely knowledge of the system state, such as full channel state information or global network status, is often infeasible due to communication overhead and physical limitations. This gives rise to a fundamental challenge we termed partial observability at scale. Traditional optimization methods, which assume full system knowledge, fail to deliver robust performance because they are either risk-oblivious, optimizing only average metrics, or risk-myopic, enforcing average constraints without guarding against high-impact tail events like severe latency spikes or QoS violations.
To address this gap, this dissertation proposes a unified framework for risk-aware intelligence in large-scale partially observable communication systems. It advances beyond conventional approaches by developing two complementary paradigms: model-based risk-aware planning and model-free risk-aware multi-agent reinforcement learning. The first paradigm is applied to antenna selection in massive MIMO with partial channel state information. It introduces a novel Risk-Aware Monte Carlo Tree Search (RA-MCTS) planner that relies on a predictive world model to forecast full channel state information from partial historical measurements. RA-MCTS plans over the resulting belief distribution of future channel states to explicitly minimize the risk of QoS violations. To improve prediction accuracy, the dissertation further develops an enhanced spatio-temporal world model featuring a Crossover Attention mechanism that modifies the standard Transformer architecture to better capture spatial and temporal correlations.
The second paradigm tackles decentralized decision-making in large-scale communication systems where accurate world models are intractable, demonstrated through asynchronous packet routing in LEO mega-constellations. This dissertation further introduces PRIMAL (Principled Risk-aware Independent Multi-Agent Learning), a model-free framework enabling satellites to make independent event-driven routing decisions. PRIMAL employs a distributional primal-dual learning method that directly constrains Conditional Value-at-Risk (CVaR) by modeling the full distribution of routing outcomes, thereby controlling worst-case risks such as extreme delays and congestion. Together, these paradigms provide principled scalable tools for managing performance risks under partial observability, laying a foundation for robust and intelligent next-generation communication systems.