Learning-based optimization of cache content in a small cell base station

Pol Blasco, Deniz Gündüz

Introduction

As mobile data traffic continues to grow, small cell base stations (sBSs) are increasingly deployed to enhance network capacity. However, these stations often face backhaul constraints, where the link to the core network becomes a bottleneck. Local caching of popular content can alleviate this pressure, but it introduces a significant challenge: base stations typically do not know content popularity in advance and can only observe requests for items already stored in their local cache. This paper proposes a learning-based framework to optimize cache content under these conditions of uncertainty.

Methodology

The authors model the cache placement problem as a Combinatorial Multi-Armed Bandit (CMAB). In this framework, each file represents an "arm," and the cache controller selects a subset of files to store in each time period.

Learning Algorithms

  • CUCB (Combinatorial Upper Confidence Bound): A baseline learning algorithm that balances exploration and exploitation by adding a confidence bonus to files with limited historical data.
  • MCUCB (Modified CUCB): The authors' primary contribution, which incorporates prior knowledge of the Zipf distribution—a common model for network traffic popularity. By assuming this structure, MCUCB can prioritize exploitation more effectively when a high degree of skewness is detected, leading to significantly faster learning.

Optimization Solver

Selecting the optimal set of files for a fixed cache capacity is a variation of the Knapsack Problem, which is NP-hard. To maintain real-time efficiency, the authors employ a Greedy Approximation (SPO-Solver). They demonstrate that in practical caching scenarios where the cache size is much larger than individual files, this greedy approach achieves over 99.9% of the optimal reward while remaining computationally lightweight.

Experimental Results

The proposed methods were evaluated through numerical simulations using Zipf-like popularity distributions and varying system parameters, including file count and user density.

Convergence and Learning Speed

A critical finding is the disparity in learning speed between standard and modified algorithms. While standard CUCB is "blind" to the underlying data structure, MCUCB leverages the Zipf assumption to reach near-optimal performance rapidly.

Long-term convergence comparison showing MCUCB reaching near-optimal rewards significantly faster than standard CUCB.

As illustrated in Figure 4, MCUCB achieves performance levels comparable to the Informed Upper Bound (IUB)—which has perfect knowledge of popularity—within approximately 5,000 iterations. In contrast, the standard CUCB algorithm remains sub-optimal even after 100,000 iterations, highlighting the value of incorporating structural priors into the learning process.

Impact of Popularity Skewness

The effectiveness of caching is highly sensitive to the skewness of content popularity, represented by the parameter $\gamma$.

Performance of various caching strategies as a function of the popularity skewness parameter $\gamma$.

Figure 1 demonstrates that as the popularity profile becomes more skewed (higher $\gamma$), the percentage of traffic fetched from the cache increases dramatically for all learning-based methods. When popularity is nearly uniform (low $\gamma$), the benefits of caching are minimal, as no small subset of files can satisfy a large portion of user requests.

Key Insights and Conclusion

  • Structural Priors: Incorporating domain-specific knowledge, such as the Zipf distribution of web traffic, can accelerate reinforcement learning by orders of magnitude.
  • Greedy Efficiency: Simple greedy algorithms are highly effective for cache management, providing near-optimal results with minimal computational overhead.
  • Observability Constraints: The system successfully learns optimal placement even with "partial sight," where it only receives feedback on files currently in the cache.

While the current model assumes a stationary popularity profile, future research could extend these learning techniques to non-stationary environments where content popularity shifts dynamically over time.

Summary updated

BibTeX
@inproceedings{blasco2014learning,
  author = {Pol Blasco and Deniz Gündüz},
  title = {{Learning-based optimization of cache content in a small cell base station}},
  booktitle = {ICC},
  year = {2014},
  url = {https://openreview.net/forum?id=Oi7tWwjHaf3},
}

Loading related content...

Figure