Adaptive Edge Caching for Non-Stationary Demand

This direction is merged into another direction: dynamic content popularity. Its last version stays readable and citable here.

Version 1 of 130 papers31 conceptsfrozen 2026-10-04

Traditional caching algorithms, such as Least Recently Used (LRU), operate on a simple premise: the recent past is a reliable predictor of the near future. In the stable environment of a core network, this heuristic performs well. However, at the "wireless edge"—the small cells and base stations closest to users—this assumption of stationarity collapses.

In modern 5G and 6G networks, content popularity is highly volatile. A video might go viral in a specific neighborhood for three hours and then vanish, or a user’s personal habits might dictate that they never watch the same content twice in a row. This research direction moves away from static caching policies toward online learning frameworks that treat caching as a dynamic estimation problem.

The Failure of the Stationary Assumption

Most classical caching models assume that content popularity follows a fixed distribution (like Zipf’s Law) that is known in advance. As Alex Dimakis and Andreas F. Molisch argue in FemtoCaching: Wireless Video Content Delivery through Distributed Caching Helpers, real-world demand is non-stationary. If a cache is optimized for yesterday's hits, it will suffer a "cache miss" crisis when today's trends shift.

To solve this, researchers have shifted from "finding the best policy" to "learning how to switch policies." Robert B. Gramacy and Scott A. Brandt pioneered this in Adaptive Caching by Refetching, using online machine learning to maintain "virtual caches." By simulating multiple algorithms simultaneously and dynamically promoting the winner to the physical cache, they allow the system to adapt to shifting workloads in real-time.

Learning Popularity in Space and Time

At the edge, popularity isn't just a global ranking; it is a "space-time" phenomenon. A file might be popular at a specific base station only during morning commute hours. Alireza Sadeghi and Georgios B. Giannakis address this in Optimal and Scalable Caching for 5G Using Reinforcement Learning of Space-Time Popularities by modeling requests as Markov processes.

By using Reinforcement Learning (RL), the cache becomes an agent that learns to prefetch content based on the expected future reward (hit rate). This is further refined by Wei Li and colleagues in A Reinforcement Learning Based Smart Cache Strategy for Cache-Aided Ultra-Dense Network, where Recurrent Neural Networks (RNNs) are used to capture the temporal dependencies of user requests, allowing the system to anticipate "bursty" traffic before it happens.

The Accuracy versus Mixing Trade-off

A critical intuition in this field is the trade-off between how accurately a system learns and how fast it can forget. In Accurate Learning or Fast Mixing? Dynamic Adaptability of Caching Algorithms, Jian Li and John C. S. Lui identify a fundamental tension:

  1. Learning Accuracy: To perfectly identify the most popular items, you need a large window of history.
  2. Mixing Time: To adapt to a sudden change (like a new viral video), you need to ignore the old history quickly.

If a cache takes too long to "mix" (reach its new equilibrium), the content trend may have already passed. This realization led to the development of algorithms like Adaptive-LRU, which tune their sensitivity to history based on the detected rate of change in the request stream.

Cooperative and Coded Adaptation

In dense networks, caches do not act in isolation. Elahe Rezaei and Babak Khalaj proposed CoM-Cache in Multi-agent Learning for Cooperative Large-scale Caching Networks, which uses Multi-Agent RL (MARL) to let neighboring caches coordinate. Instead of every cache storing the same top-10 videos, they learn to partition the content, effectively creating a larger, shared pool of storage.

This cooperation is taken a step further with coded caching. Xiongwei Wu and Jun Li demonstrate in Multi-Agent Reinforcement Learning for Cooperative Coded Caching via Homotopy Optimization that by storing fragments of mathematically "coded" files, the network can satisfy multiple different user requests with a single multicast transmission, even as the underlying demand for those files fluctuates.

Theoretical Limits of Online Caching

While much of the work is algorithmic, researchers like Rajarshi Bhattacharjee and Abhishek Sinha have sought to define the "regret" of these systems—the difference between the performance of an online learner and an omniscient observer who knew the future. In Fundamental Limits of Online Network-Caching, they prove that algorithms based on Follow-the-Perturbed-Leader (FTPL) can achieve near-optimal sub-linear regret, providing a mathematical guarantee that these adaptive systems will eventually converge to the best possible strategy, no matter how the user demand evolves.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_adaptive_edge_caching_for_non_stationary_b5821f30,
  title = {Adaptive Edge Caching for Non-Stationary Demand},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adaptive-edge-caching-for-non-stationary-demand/rdr_b5821f3095744bd9ab268c866a375413},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_adaptive_edge_caching_for_non_stationary_b5821f30_v1,
  title = {Adaptive Edge Caching for Non-Stationary Demand},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adaptive-edge-caching-for-non-stationary-demand/rdr_b5821f3095744bd9ab268c866a375413/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}