Resource Allocation for Capacity Optimization in Joint Source-Channel Coding Systems

Zhaohui Yang, Kaiyi Chi, Qianqian Yang, Zhaohui Yang 0001, Yiping Duan, Zhaoyang ZhangICC 2023 · DOI

TL;DR

This paper tackles the challenge of efficiently allocating limited wireless resources (bandwidth, power) in Joint Source-Channel Coding (JSCC) systems to maximize the number of users receiving high-quality data within their latency constraints. They propose a novel two-stage optimization approach that first minimizes power for individual user-resource block pairs and then maximizes the number of supported users, outperforming existing methods.

What Problem Does This Paper Address?

Traditional communication systems treat source coding (compressing data) and channel coding (adding redundancy for error correction) as separate steps. However, for finite block lengths (real-world scenarios), Joint Source-Channel Coding (JSCC) offers better performance by integrating these processes. Deep learning has enabled powerful JSCC models for various data types like images and text.

Despite these advances, a critical bottleneck remains: resource allocation. Base stations (BS) have limited power and bandwidth to serve multiple users, each with unique quality-of-service (QoS) requirements (e.g., desired image quality, maximum acceptable delay). Existing JSCC resource allocation methods often focus on maximizing overall system performance metrics, but they don't adequately address the problem of maximizing the number of users that can be successfully served, especially when resources are scarce. This paper aims to fill that gap by developing a resource allocation strategy that prioritizes serving the most users possible while strictly adhering to their latency and quality demands.

Methods

The authors propose a two-stage optimization framework to address the complex mixed-integer non-linear programming problem of resource allocation in JSCC systems.

  1. Problem Decoupling: The original problem is decomposed into two subproblems without loss of optimality. This is a key methodological contribution.
  2. Subproblem 1: Power Minimization (P2): For each potential user-resource block (RB) pair, this subproblem determines the minimum transmission power required to satisfy a user's latency and performance (PSNR) constraints. This involves optimizing the compression ratio (CR) and power allocation. They achieve this by first finding the maximum feasible CR given the latency constraint and then using a bisection method to find the minimum SNR required for the desired PSNR.
  3. Subproblem 2: System Capacity Maximization (P3): Once the minimum power for each user-RB pair is known, this subproblem focuses on selecting the optimal user-RB pairings to maximize the total number of supported users, subject to the BS's total power budget. This is formulated as a binary integer programming problem.
  4. Solving Subproblem 2: To efficiently solve the user-RB pairing problem, they relax the binary variables to continuous ones and solve the resulting linear program using Karmarkar's algorithm. This approach is proven to yield the same optimal solution as the integer problem.
  5. Bisection Search for User Count: To find the maximum number of users (J) that can be supported within the power budget, they employ a bisection search over J, iteratively solving the relaxed linear program (P5) for different values of J.

Datasets

The paper uses the CIFAR10 dataset for evaluating image transmission. This dataset consists of 60,000 images of size 32x32 pixels. The JSCC model used in the experiments was pre-trained on this dataset. No new benchmarks or datasets were introduced.

Key Insights and Ideas

  • Adaptive Compression is Crucial: The paper strongly advocates for adaptive compression ratios (CR) per user, demonstrating that fixed CR schemes lead to significant performance degradation. This is because a fixed CR might be too high for some users (requiring excessive power for high quality) or too low for others (failing to meet quality constraints).
  • Decoupling Simplifies Complexity: Decomposing the complex resource allocation problem into power minimization and then capacity maximization significantly simplifies the solution process, making it computationally tractable.
  • Power Budget is the Bottleneck: The simulations highlight that when the BS power budget is limited, the proposed method effectively maximizes the number of served users, outperforming methods that might be optimal when power is abundant (like the Hungarian-based approach).
  • Latency and Quality are Intertwined with Power: The results clearly show how latency and PSNR requirements directly translate into power demands. Users with stricter latency or higher PSNR needs require more power, and the BS's limited power dictates how many such users can be supported.

Key Results

  • The proposed resource allocation method consistently outperforms benchmark methods (Hungarian Based, Random Pairing, Uniform Power) in terms of the number of supported users across various BS power levels.
  • The performance gap between the proposed method and the optimal Hungarian-based method is small when BS power is abundant, but the proposed method significantly outperforms it under limited power.
  • Using adaptive compression ratios leads to a substantial increase in the number of supported users compared to fixed CR schemes. For instance, a fixed CR of 1/24 showed particularly poor performance.
  • The access ratio (proportion of supported users) is directly impacted by delay constraints; users with tighter delay requirements (e.g., 4ms) are harder to support due to higher power demands. Similarly, higher PSNR requirements also reduce the access ratio.

Limitations and Future Work

  • Computational Complexity: While the proposed algorithm is polynomial, the complexity of solving the linear program (P5) using Karmarkar's algorithm can still be significant for very large numbers of users and RBs. The paper mentions this complexity in Section V.D.
  • Approximation of PSNR-SNR Relationship: The paper uses a polynomial function to approximate the relationship between PSNR, CR, and SNR. The accuracy of this approximation could impact the overall performance.
  • Future Work: The paper doesn't explicitly detail future work directions, but potential avenues could include extending this framework to more complex network topologies, incorporating dynamic channel conditions, or exploring different QoS metrics beyond PSNR and latency. The adaptive compression aspect itself suggests further research into more sophisticated CR selection strategies.

Summary updated

BibTeX
@inproceedings{yang2023resource,
  author = {Zhaohui Yang and Kaiyi Chi and Qianqian Yang and Zhaohui Yang and Yiping Duan and Zhaoyang Zhang},
  title = {{Resource Allocation for Capacity Optimization in Joint Source-Channel Coding Systems}},
  booktitle = {ICC},
  year = {2023},
  doi = {10.1109/icc45041.2023.10279466},
}
Figure