Source-linked AI summary
Energy-Efficient Radio Resource Allocation for Federated Edge Learning
Qunsong Zeng, Yuqing Du, Kin K. Leung, Kaibin Huang
TL;DR
The paper studies how to reduce edge-device energy consumption in federated edge learning without compromising learning performance. It develops energy-efficient radio resource management strategies for joint bandwidth allocation and scheduling, adapting to channel states and computation capacities. The resulting policies allocate more bandwidth to bottleneck devices, while scheduling priorities favor devices with better channels and computation capacities.
Problem
Energy-efficient communication for FEEL had not been explored, although model training and transmission consume energy and edge devices have limited battery lives.
Method
The paper proposes energy-efficient RRM strategies that jointly optimize bandwidth allocation and device scheduling under a learning-speed constraint.
Results
The derived policies allocate more bandwidth to devices with weaker channels or poorer computation capacities, while scheduling priorities favor devices with better channels and computation capacities.
Takeaways & Limitations
Energy-efficient RRM can reduce FEEL device energy consumption while warranting learning performance.
Abstract
from arXiv · showhide
Edge machine learning involves the development of learning algorithms at the network edge to leverage massive distributed data and computation resources. Among others, the framework of federated edge learning (FEEL) is particularly promising for its data-privacy preservation. FEEL coordinates global model training at a server and local model training at edge devices over wireless links. In this work, we explore the new direction of energy-efficient radio resource management (RRM) for FEEL. To reduce devices' energy consumption, we propose energy-efficient strategies for bandwidth allocation and scheduling. They adapt to devices' channel states and computation capacities so as to reduce their sum energy consumption while warranting learning performance. In contrast with the traditional rate-maximization designs, the derived optimal policies allocate more bandwidth to those scheduled devices with weaker channels or poorer computation capacities, which are the bottlenecks of synchronized model updates in FEEL. On the other hand, the scheduling priority function derived in closed form gives preferences to devices with better channels and computation capacities. Substantial energy reduction contributed by the proposed strategies is demonstrated in learning experiments.
I. INTRODUCTION
The paper addresses energy-efficient radio resource management for federated edge learning, where prior communication designs largely overlooked device energy consumption. It proposes joint bandwidth allocation and scheduling strategies that reduce total device energy while preserving learning performance.
- Motivation: FEEL coordinates server-side global updates with local models computed from edge-device data, avoiding data uploading to preserve privacy.
- Contributions: The work proposes joint bandwidth allocation and scheduling to minimize total device energy under a learning-speed constraint.
- Motivation: Existing FEEL communication schemes focus on learning acceleration, while energy consumption remains out of scope despite devices’ limited battery lives.
- Contributions: For a fixed device set, the closed-form bandwidth policy uses all allowed uploading time and allocates less bandwidth to devices with stronger computation and better channels.
- Contributions: The scheduling strategy alternates scheduling and bandwidth-allocation subproblems, with closed-form priorities favoring devices with better computation capacities and channels.
II. SYSTEM MODEL
The FEEL system alternates server aggregation of uploaded local models with broadcasting of the global model. The server uses model, channel, and computation information to manage each communication round.
- The system consists of one edge server and K edge devices participating in iterative communication rounds.
- FEEL alternates between aggregating transmitted local models into a global model and broadcasting that global model back to devices.
- The server is assumed to know the model size, multiuser channel gains, and local computation capacities through feedback.
- For each communication round, the server determines an energy-efficient scheduling and bandwidth-allocation strategy.
A. Multiple-access Model
The multiple-access model uses OFDMA for uploading local models under synchronous timing constraints. Bandwidth is divided among selected devices according to channel and timing variables.
- Local model uploading uses orthogonal frequency-division multiple access (OFDMA).
- Device k receives bandwidth γ_kB, where γ_k is its allocation ratio and B is the total system bandwidth.
- Synchronous updates impose a per-round time constraint covering local computation and model uploading.
- Heterogeneous computation capacities are represented by differences in devices’ local computation times.
- The model-uploading time is denoted t_k, and the notation is re-denoted consistently for subsequent analysis.
B. Energy Consumption Model
The energy model separates transmission energy for uploading local models from local-training energy. Learning speed is represented by the number of scheduled devices, which is constrained by finite radio resources.
- Each device’s per-round energy consumption has transmission and local-model-training components.
- The achievable upload rate depends on transmission power, channel gain, bandwidth, and channel-noise variance.
- The scheduling indicator β_k equals 1 when device k uploads and 0 otherwise.
- All devices have identical local-training energy E_comp because they train the same model with the same backpropagation algorithm.
- The number of scheduled devices measures learning speed, while finite radio resources limit how many devices can be scheduled.
III. ENERGY-EFFICIENT BANDWIDTH ALLOCATION
For active devices, the paper minimizes total energy by jointly choosing transmission times and bandwidth allocations under synchronous timing constraints. The resulting policy prioritizes bottleneck devices for bandwidth while contrasting with rate-centric allocation.
- The RRM problem minimizes total energy while determining bandwidth partitioning and uploading times for active devices.
- The objective is non-increasing in each device’s transmission time and bandwidth allocation.
- The optimal bandwidth policy is characterized in Theorem 1 through a solved Lagrange multiplier and restricted transmission times.
- More bandwidth is allocated to devices with poorer computation capacities because they bottleneck synchronized updates and need shorter uploading times.
- More bandwidth is allocated to devices with weaker channels, favoring bandwidth over increased transmission power for energy minimization.
- Unlike rate-maximization methods that favor stronger channels, the proposed learning-centric policy favors weaker channels and poorer computation capacities.
IV. ENERGY-AND-LEARNING AWARE SCHEDULING
The section develops an energy-and-learning-aware scheduling approach that jointly allocates bandwidth and selects uploading devices under coupled constraints. It uses relaxation-and-rounding and alternates bandwidth allocation with scheduling to obtain an efficient solution.
- Selecting a subset of devices can reduce sum energy consumption while meeting learning-speed requirements.
- The joint problem is non-convex because device-selection indicators are integer constrained.
- Relaxation-and-rounding replaces binary selection indicators with continuous values before rounding the solution to integers.
- The proposed iterative method alternates bandwidth allocation and scheduling subproblems until convergence.
- The closed-form selection priority favors devices with high computation capacity and good channels.
- Selection priority scales linearly with allowed transmission time and approximately with log(h_k), making transmission time or computation capacity dominant over channel gain.
V. SIMULATION RESULTS
The simulations evaluate the proposed resource-management strategies in a 50-device OFDMA FEEL system using MNIST and a convolutional neural network. The setup specifies wireless, computation, model-size, and dataset parameters.
- The simulation uses K = 50 edge devices and an OFDMA bandwidth of B = 1 MHz.
- Channel gains follow independent Rayleigh fading with average path loss 10^-4, while channel-noise variance is N_0 = 10^-8 W.
- The learning task classifies handwritten digits from MNIST, with each device randomly assigned 20 samples and model size L = 10^4 bits.
- The model is a 6-layer CNN with two 5 × 5 convolution layers, ReLU activations, 2 × 2 max pooling, a 50-unit fully connected layer, and softmax output.
1) Energy-efficient bandwidth allocation:
The simulations compare the proposed bandwidth allocation and scheduling policies with uniform allocation and baseline selection as communication-round time varies. Energy decreases with more transmission time, while learning accuracy improves as more devices can upload.
- The proposed RRM policy is benchmarked against uniform bandwidth allocation, which assigns equal bandwidth to edge devices.
- Total energy consumption decreases as the allowed communication-round time T increases for both policies.
- The proposed optimal policy outperforms the baseline scheme in total energy consumption.
- Energy-and-learning aware scheduling: The baseline learning performance is independent of T because its number of scheduled uploading devices remains fixed at K = 50.
- Energy-and-learning aware scheduling: The proposed algorithm’s average learning accuracy increases with T and approaches the baseline for large T.
- Energy-and-learning aware scheduling: 70% to 98%: the energy reduction ratio approximately spans this range and decreases as T grows.
VI. CONCLUDING REMARKS
The paper proposes energy-efficient FEEL radio-resource management that adapts bandwidth allocation and scheduling to channel states and computation capacities. Its strategies reduce sum device energy consumption while providing a learning-speed guarantee.
- The proposed RRM jointly addresses bandwidth allocation and scheduling for federated edge learning.
- The strategies adapt to channel states and computation capacities to reduce sum device energy consumption.
- The work presents an initial exploration of energy-efficient RRM for federated edge learning.
- Future extensions include asynchronous model updates, update sparsity, local-computing energy effects, and local batch-size adaptation.
APPENDIX
The appendix derives closed-form optimal solutions using convexity, KKT conditions, and the Lambert W function. It also establishes how the optimal allocation varies with channel gains and computation-related parameters.
- Proof strategy: Convexity and KKT conditions are used to derive the optimal allocation solution.The derivation introduces multipliers for inequality and equality constraints before solving the resulting conditions.
- Closed-form solution: The closed-form solution uses the Lambert W function, with ν⋆ determined from the equality constraint.The derivation explicitly identifies W(·) as the Lambert W function.
- Conclusion: The appendix completes the proof after establishing the derived allocation and its monotonicity properties.The supplied proof passages explicitly mark completion after the derivations.
- Monotonicity: γ⋆_k is non-increasing with respect to h^2_k and non-decreasing with respect to T_k.These monotonicity properties follow from the behavior of W(x), x(T_k), and the intermediate expressions.
C. Proof of Theorem 2
The proof treats the optimization as a convex problem and characterizes the constrained optimum by projecting an unconstrained value onto the interval [0,1].
- Case analysis: If ˆβ_k is below 0, the optimum is β_k = 0; if it lies in [0,1], the optimum is β_k = ˆβ_k.These are the first two cases in the constraint analysis.
- Optimal point: If ˆβ_k exceeds 1, the optimum is β_k = 1, yielding the summarized clipped optimal point.The three cases together define the constrained optimum.