NOMA power allocation moves to per-user learning: fairness index 0.96 and outage below 10% at 10 dB in a three-user simulation
Synopsis
In downlink NOMA, power allocation has relied on fixed, equal, or centralized learned rules that struggle to serve weak-channel users and throughput at once. A Kampala International University team models each user as an independent learning agent, coordinates them with a shared reward over local observations, and builds imperfect-SIC residual interference into the learning environment. In a three-user simulation the scheme reached about 18 bps/Hz and a fairness index of 0.96 at 30 dB SNR, with outage below 10% at 10 dB.
Interpretation
Power allocation shifts from a base-station decision to one decision per user, so weak-channel users no longer receive only a preset share. Fixed and equal allocation follow preset rules, while the genetic algorithm and centralized deep Q-network let a single decision-maker hold the global state. The comparison runs in one three-user downlink NOMA simulation, with all five schemes under the same Monte Carlo setup averaged over 10,000 iterations.
Each user observes only its own channel gain, experienced interference, and previous power, picks a power coefficient, and normalization by the sum of coefficients enforces the total power constraint. Decisions come from independent deep Q-networks whose agents share one global reward combining sum rate, transmit power, and outage. Training is fixed at 500 episodes, learning rate 0.001, and discount factor 0.95, over Rayleigh fading with QPSK modulation.
At 30 dB SNR the scheme reaches about 18 bps/Hz with a fairness index of 0.96, keeps outage below 10% at 10 dB, and drives BER below 10 to the minus 3 at high SNR. The centralized deep Q-network holds a slightly higher sum rate at high SNR but trails on fairness, outage, and BER. These figures come from a three-user simulation with 1 W total power and 10 to the minus 9 W noise, repeated under imperfect channel state information.
Perspective
The conclusions rest on a three-user downlink NOMA simulation with total transmit power normalized to 1 W, noise power of 10 to the minus 9 W, Rayleigh fading, and QPSK modulation. Evaluation covers sum rate, outage probability, fairness index, and BER, and is repeated under imperfect channel state information. The authors propose next steps toward high-mobility and large-scale networks, actor-critic, federated and meta-learning paradigms, and real-time implementation with field trials on a software-defined radio platform.
The authors note that the centralized deep Q-network keeps a slightly higher sum rate at high SNR, so the decentralized scheme trades some throughput for fairness, outage, and BER. Under imperfect channel state information all schemes lose sum rate; the decentralized one leads at low SNR but is overtaken by conventional schemes at high SNR. Training complexity, convergence speed, and generalization to unseen channel conditions are named as open issues for deployment.