A hybrid resource-allocation method recorded the highest reported computational throughput in a simulated MapReduce system using heterogeneous wireless devices powered by renewable energy harvesting. Called DDPG-CVX, it pairs a learning-based decision component with a convex optimization step for the remaining per-slot resources. The numerical evaluation modeled five devices.
The primary measure was average system computational throughput, the study's measure of aggregate computing load over time. DDPG-CVX reached 1.25 times the throughput of DDPG-only, 1.51 times that of Greedy Allocation (GA), 1.75 times that of Random Budget (RB), 2.87 times that of Average Allocation (AA) and 32.36 times that of MAX Frequency (MAXF).
Because the endpoint is a system average, the ratios describe the relative ranking of the algorithms under the same modeled conditions rather than the output of one device in isolation.
The calculation behind the comparison
At the center of the paper is a long-run constrained optimization. It jointly handles computing load, phase-time allocations, transmit power and per-device energy consumption, while respecting battery-evolution, CPU-frequency and latency constraints.
DDPG-CVX is designed for online, slot-by-slot use without prior channel-state or energy-arrival knowledge. DDPG selects an energy budget for each device. An embedded convex solver then chooses the remaining variables and resolves that per-slot problem to global optimality.
In practical terms, the method divides the decision between setting an energy budget and completing the constrained allocation. The learning component handles the budget choice, while the solver resolves the rest of the slot.
Inside the simulated system
The numerical evaluation used N = 5 heterogeneous wireless devices, 30 MHz of bandwidth and 100 ms time slots. Channel gains were complex Gaussian, and harvested energy was generated independently for each device and time slot from a bounded-uniform process.
The DDPG actor and critic networks each had two hidden fully connected layers, with 32 neurons in the first and 64 in the second, using ReLU activations.
The ranking held across the tests
Training results favored the hybrid method as well. DDPG-CVX converged to the highest throughput within 1,000 episodes, while DDPG-only converged more slowly and to a lower value.
Across the tested range of maximum battery capacity, Emax, from 100 to 10,000 µJ, DDPG-CVX achieved the highest throughput throughout.
When the model's capacitance coefficient, kn, was scaled from 0.4 to 2.4, throughput declined for every algorithm. DDPG-CVX nevertheless remained highest across the full tested range.
The results were sensitive to training choices. Throughput changed with the discount factor and the actor and critic learning rates; among the settings tested, the highest throughput came with γ = 0.95 and ηa = ηc = 2 × 10−3.
The result remains tied to the model
The result is tied to the numerical conditions reported in the paper: five devices, fixed bandwidth and slot duration, complex Gaussian channel gains and independently generated bounded-uniform energy arrivals. It shows how the tested algorithms ranked in that setup, rather than establishing performance across other hardware or deployment conditions.
The document is an arXiv preprint, identified as arXiv:2608.25549v1 [eess.SY] and dated 26 August 2026.
The work was supported by the Shanghai Municipal Science and Technology Commission Foundation grant 25DP1500300.
Paper data and sources
Original title: Throughput Maximization for MapReduce-Based Collaborative Computing over Energy-Harvesting Wireless Devices
Authors: Yuhang Li, Siqi Sun, Hongen Zheng et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text