Finding structure beneath complex systems.

Random access is one of the two fundamental forms of multiple access and is widely adopted across modern communication systems, including Wi-Fi, cellular, and satellite networks, among many others. Yet its theory has historically been fragmented: models were tailored to particular protocols, traffic assumptions, or performance measures, sometimes producing findings that were difficult to reconcile.

Our unified analytical framework provides a common basis for analyzing and optimizing all random access schemes, regardless of the performance metric adopted. Throughput, delay, information rate, stability, and other objectives can all be studied within the same framework, allowing results across different protocols and metrics to be understood and compared coherently.

Distributed users must decide when to transmit using limited local information. Reinforcement learning offers a new route to designing access strategies that adapt through interaction rather than relying on a prescribed protocol.

Our work pursues two complementary directions. First, we apply reinforcement learning to discover distributed access strategies for fundamental throughput, delay, and fairness objectives. Second, we use the unified theory of random access to identify the strategies learned, explain their behavior, and guide their design and parameter selection. The goal is to make learning-based access not only effective, but analytically understandable.

Classical channel capacity characterizes the maximum rate at which information can be transmitted reliably when arbitrarily long coding and transmission delays are allowed. As an asymptotic measure, it does not capture the fundamental tradeoff between information rate and delay that is central to low-latency and time-sensitive communication.

By modeling the information-transmission process as a queue of codewords, our work incorporates access and queueing delays into the characterization of information delivery. This leads to the maximum information output rate under a delay constraint and, more recently, under information aging—where the amount of successfully delivered information depends on how long delivery takes.

The cellular structure, used since the first generation of mobile networks, decomposes a network according to base-station coverage and assigns each user to one cell. As base stations become denser, a user may be close to several of them but associate with only one, leading to strong inter-cell interference.

Looking beyond the cellular structure, fully cell-free networking allows all base stations to jointly process signals from all users, but it is not scalable as the densities of both base stations and users increase. A scalable architecture must divide a large network into parallel subnetworks; if that decomposition is poorly designed, however, strong interference between subnetworks can severely degrade overall performance.

We therefore propose optimal network decomposition: forming as many parallel subnetworks as possible while keeping inter-subnetwork interference limited. We formulate this problem as a bipartite graph partitioning problem that jointly groups users and base stations into subnetworks. Our subsequent work incorporates rate and joint-processing constraints into this framework for practical clustered cell-free networking.