cs.LGOct 3, 2026

Multi-Agent Spectrum Sharing

Authors: Job Elliott, Graduate Student Member, IEEE, Justin G. Metcalf, Golnaz Habibi

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

This project explores how multiple cognitive radars can learn to share limited wireless spectrum with other radio users without interfering with one another. Using machine learning (ML), each device independently decides where and how widely to transmit within a fixed 100 MHz band. The system analyzes real or simulated signal activity to detect which parts of the spectrum are currently in use and which are open. Based on this information, the devices adapt their transmission choices to avoid crowded frequencies while making efficient use of available space. The goal is to develop a flexible, scalable approach to spectrum sharing that could support future wireless communication systems. Experimental results using both over-the-air software-defined radio (SDR) recordings and simulated environments demonstrate that the proposed meta-learning approach consistently balances competing objectives better than conventional reinforcement learning (RL) methods in multi-agent spectrum-sharing scenarios. Across five multi-agent benchmark environments, our proposed method achieved the highest average reward among the primary baseline algorithms while simultaneously maintaining low collision rates and stable transmission behavior.

Figures & tables

Fig. 1 : M-FOS Architecture
Fig. 2 : Collected 0.02 Second Spectrum Data Samples
TypeBW (MHz)Tx / IdleAction Update
Skinny1.46–2.932.3–3.58 ms / 19.46–21.50 msEvery 307.2 ms
Wide14.65–17.091.02–2.56 ms / 92.16–102.40 msEvery 1024 ms
Pulsed2.44–4.880.21–0.51 ms / 0.51–0.82 ms (within 20.48–51.2 ms burst cycles)Every 1024 ms
Rectangular5.86–66.410.46–0.56 ms / 0.1 ms–2.56 msEvery Pulse
TABLE I : Simulation emitter parameters.
Fig. 3 : Simulation Data .02 Second Sample
TABLE II : Performance comparison of all methods. Bold font indicates best value for that column while underlined numbers indicate the best value among the cognitive agents excluding the Random Static and SAA agents. Each value represents the average value per agent across three independent runs.
TABLE III : Performance comparison of all methods. Each value represents the average of three values per agent type.
Fig. 4 : Multi-Agent Spectrum Occupancy between PPO, M-FOS, and DQN. First, middle, and last 20,000 Samples, from bottom to top respectively.
Fig. 5 : Comparison of average reward performance.
Fig. 6 : Pretrained model comparison of average reward performance in the multi-agent environment using ARRC 2.4–2.5 GHz dataset.
Fig. 7 : Pretrained model comparison of average reward performance in the multi-agent environment using Union 2.4–2.5 GHz dataset.
AgentΔBW↓\Delta BW\downarrowσΔBW↓\sigma_{\Delta BW}\downarrowΔCF↓\Delta CF\downarrowσΔCF↓\sigma_{\Delta CF}\downarrow
SAA17.136614.863215.788616.0901
PPO2.72545.17411.397 22.5187
DQN0.23190.86409.557023.6005
M-FOS0.22930.85421.68403.1335
TABLE IV : Multi-Agent Distortion Statistics 2.4–2.5 GHz ARRC Data
Fig. 8 : Range Doppler Plots
Agents2.4–2.5 GHz ↑\uparrow2.59–2.69 GHz ↑\uparrowSimulation ↑\uparrow
Random Static-7,713-15,196-4,386
SAA-125,317-133,174-217,405
PPO-6,603-10,690-27,498
DQN-11,894-14,001-6,495
M-FOS-1,213-9,599-1,049
TABLE V : Multi-Agent Total Rewards Across Frequency Bands (ARRC and simulation datasets)
Figure 18
Figure 19
Figure 20
Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

ParameterValue
Time horizon1024
Discount factor ( γ\gamma )0.95
GAE lambda0.95
Clipping parameter ( ϵ\epsilon )0.2
Optimization epochs10
Learning rate5×10−45\times 10^{-4}
Appendix
TABLE VI : PPO Hyperparameters
ParameterValue
Replay memory size2000
Batch size32
Discount factor ( γ\gamma )0.9
Learning rate1×10−31\times 10^{-3}
Target network update interval1000 pulses
Hidden layer sizes[512, 256, 84]
Appendix
TABLE VII : DQN Hyperparameters
ParameterValue
Population size5
Mutation scale0.05
Elite fraction0.4
Fresh population fraction0.2
Appendix
TABLE VIII : M-FOS Hyperparameters
ParameterValue
Clean transmission weight1.0
Collision penalty weight30.3
Bandwidth distortion penalty weight0.5
Center-frequency distortion penalty weight0.5
Appendix
TABLE IX : Reward Function Parameters
More in this topic

Wireless Resource Allocation

Wireless resource allocation assigns spectrum, power, beams, and computing capacity to meet throughput, latency, fairness, and energy goals.

See all 175 papers

Explore similar work

CardsList

  1. SpecMind: Enabling Spectrum Intelligence via Multi-Agent Hybrid Retrieval-Augmented Generation

    Aug 31, 2026Songwei Dong, Bingyan Lu, Makayla Kienlen +2Retrieval-Augmented GenerationWireless Communications

    Department of Electrical and Computer Engineering, University of Virginia · Department of Electrical Engineering, University of Notre Dame · Baskin School of Engineering, University of California, Santa Cruz

  2. WiSDoM: Wireless Sparse Decision Transformer with Mixture-of-Experts for Multi-Task Mobile Network Optimization

    Aug 31, 2026Fatih Temiz, Shavbo Salehi, Melike Erol-KantarciSparse Mixture-of-ExpertsDecision Transformer

    School of Electrical Engineering and Computer Science, University of Ottawa, Ottawa, Canada

  3. MUTE: Return-Preserving Communication Unlearning for Efficient Multi-Agent Coordination

    Jul 3, 2026Rui Zuo, Qinwei Huang, Mingyang Li +3Multi-Agent System OptimizationMulti-Agent Coordination

    Syracuse University · Air Force Research Laboratory