RISK-AWARE INTELLIGENCE FOR LARGE-SCALE PARTIALLY OBSERVABLE COMMUNICATION SYSTEMS

The evolution towards Sixth-Generation (6G) communication systems is characterized by an unprecedented increase in scale, decentralization, and dynamism. This trend is evident by terrestrial networks employing massive Multiple-Input Multiple-Output (MIMO) antenna arrays and non-terrestrial networks composed of ultra-dense Low Earth Orbit (LEO) satellite constellations. In such large-scale environments, obtaining a complete and timely view of the true system state, e.g., complete channel state information (CSI) or global network state, is often infeasible due to prohibitive communication overhead and physical constraints. This gives rise to a fundamental challenge we term partial observability at scale. Traditional control and optimization methods, which typically assume complete system knowledge, fail to provide robust solutions in such environments. Crucially, these methods fail because they do not adequately handle the risks of partial observability. Many frameworks are risk-oblivious, optimizing for average performance while ignoring critical QoS degradation caused by incomplete state knowledge. Even recent constrained approaches are risk-myopic, focusing on average performance constraints while failing to control for high-impact tailend events like severe latency spikes or QoS breaches. This leaves the system vulnerable to unacceptable performance violations. Recognizing this gap, this dissertation argues that there is an urgent need for intelligent and autonomous decision-making frameworks capable of operating reliably under partial observability at scale and managing the associated risks effectively. This leads to the central research problem addressed herein: How can communication agents make robust and risk-aware decisions in large-scale partially observable communication systems? To answer this question, this dissertation moves beyond risk-oblivious and risk-myopic approaches by proposing a unified framework for risk-aware intelligence in large-scale partially observable communication systems. We develop this framework through two complementary paradigms: model-based risk-aware planning and model-free risk-aware reinforcement learning, demonstrated on antenna selection in massive MIMO with partial CSI and asynchronous packet routing in LEO mega-constellations, respectively.