In power-domain non-orthogonal multiple access (NOMA), a base station sends to several users on the same frequency at the same time. It separates them only by how much transmit power each one gets. The weak, far user gets most of the power. The strong, near user decodes and subtracts the weak user's signal, a step called successive interference cancellation (SIC), and then decodes its own. How the power is split decides who gets what. More power to the strong user raises total throughput. More to the weak user raises fairness. Less power overall saves energy. Operators want to move along this three-way trade-off as conditions change, and they need an answer for every fresh channel measurement within milliseconds.
This note proposes a single neural network to do that. It takes two inputs: the current channel estimates and an operator preference vector that weights sum rate, worst-user rate and transmit power. It outputs a power allocation on the constrained Pareto front, meaning the set of best achievable trade-offs that respect every user's minimum-rate requirement and the power budget. The key ingredient is the network's output layer. It is a closed-form map, derived from NOMA's own decoding recursion, that turns arbitrary network outputs into allocations that satisfy every rate and power constraint exactly whenever the instance is feasible at all. Constraints are never learned approximately through penalties. The question is whether this structure lets an amortized model match per-instance solvers on solution quality at a tiny fraction of their latency, and whether it beats the obvious alternative of a preference-conditioned network trained with constraint penalties.
Core Claim
Downlink NOMA has a sharp feasibility cliff. On the Performance of Non-Orthogonal Multiple Access in 5G Systems with Randomly Deployed Users analyzes random deployments with fading and shows a "critical condition": if the power coefficients and target rates are mismatched, outage probability is 1 at every SNR. Penalty-trained networks sit right on this cliff, so they will put a small but real fraction of allocations on the wrong side of it.
The structure that creates the cliff also removes it. Write the noise-normalized channel gains as $g_1 \le \dots \le g_K$, the powers as $p_k$, and the per-user rate targets as $r_k$, with $\epsilon_k = 2^{r_k}-1$. Under perfect channel knowledge and the standard decoding order, user $k$ meets its target exactly when
$$p_k \ge \epsilon_k \Big(\sum_{j>k} p_j + 1/g_k\Big).$$
Applying this recursively from the strongest user down gives a feasible allocation. We parameterize every power as its required minimum plus a nonnegative surplus $s_k$. Total power is then affine in $s$, so a surplus direction on the simplex and a budget fraction, both produced by the network, can be rescaled in closed form so that total power never exceeds the budget $P_{\max}$.
Every output of the network is therefore feasible. Learning only has to choose where on the feasible set to land for a given preference. Training is unsupervised: the loss is a smooth Tchebycheff scalarization of the three objectives, the aggregation used by Dealing With Structure Constraints in Evolutionary Pareto Set Learning, averaged over sampled channels and preferences.
We are candid about where this idea is trivial. In the single-carrier, perfect-knowledge case with sum rate alone, the optimum is close to closed form: give all surplus to the strongest user. The proposal earns its keep in three harder settings, and the experiments are built around them. The first has three objectives, where the front is a genuine two-dimensional surface rather than a curve. The second is multi-carrier operation, where each user's power is also split across subcarriers. The third has imperfect channel estimates, where decoding feasibility is no longer automatic and the rate constraints become chance constraints.
Research Program
The first component is the feasibility map itself. We need its multi-carrier extension, with a per-carrier recursion and a shared budget split across carriers through a second simplex output. We also need its imperfect-knowledge extension. There, each $g_k$ in the recursion is replaced by a lower quantile of the posterior over the true gain, given the estimate and the estimation-error variance. For each strong user, an explicit check confirms that the SIC decoding condition still holds at that quantile. The quantile level is tuned on held-out channels until the empirical outage rate meets the target. This converts the chance constraint into a deterministic backoff.
When an instance is infeasible, meaning the minimum required power already exceeds $P_{\max}$, the map returns the allocation that minimizes total constraint violation and flags the instance. Here we borrow the survey's point that treating violation as a quantity to minimize, rather than as a hard filter, avoids discarding useful near-feasible configurations (A Survey on Evolutionary Constrained Multiobjective Optimization; Immune Clonal Multi-Objective Optimization Algorithm for Constrained Optimization). User clustering and subcarrier assignment are fixed by a standard channel-gain pairing heuristic. This is deliberate, so that the contribution is isolated to continuous power allocation. The joint discrete problem is left out of scope.
The second component is the network. It is a preference-conditioned multilayer network in the spirit of Learning the Pareto Front with Hypernetworks, which generates a whole family of trade-off models from one preference input. Unlike that work, ours is also conditioned on a per-instance channel state, so it learns a different front for each channel draw rather than one front for a fixed dataset. A Hyper-Transformer model for Controllable Pareto Front Learning with Split Feasibility Constraints handles fragmented feasible regions with attention and expert mixtures. We test a simple version of that architecture only if the multi-carrier fronts turn out to be disconnected.
The third component is analysis. We measure how far the learned front sits from per-instance optima. We also check whether the network concentrates errors on weakly Pareto-optimal boundary segments, the failure mode Dealing With Structure Constraints in Evolutionary Pareto Set Learning warns about for continuous Pareto set learning. Finally, we measure how the backoff trades front quality for outage reliability as the channel-estimation error grows.
This proposal is clearly separate from the existing Pareto work on NOMA. Beam Design and User Scheduling for Non-Orthogonal Multiple Access with Multiple Antennas Based on Pareto-Optimality slides a single "rate knob" along a two-user boundary under perfect knowledge. Optimal Throughput Fairness Tradeoffs for Downlink Non-Orthogonal Multiple Access Over Fading Channels solves one two-user ε-constraint problem per fairness target. minPIC: Optimal Power Allocation in Multi-User Interference Channels minimizes power for fixed rate targets with 2–3 users on one subcarrier. All three motivate this proposal because they compute one point, or one curve, per solve. Constrained Multi-Objective Reinforcement Learning with Max-Min Criterion supplies a related idea, dual-learned fairness weights under budget constraints. It differs in three ways: it produces one max-min policy, it enforces constraints only in expectation over trajectories, and it has never been applied to wireless problems.
Baselines
The naive reference is fractional transmit power allocation, which sets power proportional to a power of the inverse channel gain. Energy-Efficient Resource Allocation in NOMA Heterogeneous Networks reports that it achieves 27% higher energy efficiency than equal-power OFDM in a 20-small-cell network with imperfect channel knowledge. We sweep its exponent and a power-scaling factor so that it also traces a front, and we score that front on the same test channels. Equal power and plain orthogonal access are included as floors.
The strongest competitor is a per-instance solver run to convergence for each channel draw and each of a dense grid of preferences. For $K \le 3$, we use convex reformulation or fine grid search of the surplus variables, which gives near-exact fronts. For larger $K$ and multiple carriers, we use a constrained evolutionary solver with a large evaluation budget: CCMO with the operator selection of Constrained Multi-objective Optimization with Deep Reinforcement Learning Assisted Operator Selection, which ranks best across 42 constrained benchmark problems. This baseline defines the reference front. On two-user channels we also check it against the closed-form policies of Xing et al.
The only reported numbers we can anchor to are these. minPIC reports 20–40% transmit-power savings over orthogonal access "in some cases" in 2–3 user interference channels. Deep Learning Based Radio Resource Management in NOMA Networks: User Association, Subchannel and Power Allocation reports that its learned power allocation reaches about 98% of the energy efficiency of gradient iteration. No reported hypervolume result was found on Lacuna for any NOMA power-allocation method. Beating the solver means coming within 2% of its hypervolume, the volume of objective space the front dominates, with zero constraint violations, at inference latency below a 1 ms coherence-time budget where the solver needs orders of magnitude longer. We also require the power-only corner of our front to reproduce savings over orthogonal access of the size minPIC reports in the 2–3 user setting.
The key ablation is the same network, data and tuning budget with the feasibility map replaced by a softplus output and a quadratic violation penalty with tuned weight, as in standard Pareto set learning. A second ablation keeps the map but drops the channel quantile backoff. All learned variants get the same channel samples, the same number of gradient steps and the same hyperparameter-search budget.
Why not just…?
Why not just run a convex or SCA solver per preference at decision time? For small single-carrier instances, this is a real threat. If per-instance optimization fits within the coherence time, amortization buys little, and we will report that honestly. The experiment that settles it is a latency-versus-hypervolume curve as $K$ and the number of carriers grow. minPIC itself stops at 3 users on one subcarrier and needs an outer binary search, which suggests per-instance cost rises quickly, but we will measure this rather than assume it.
Why not just train a preference-conditioned network with a constraint penalty? This is the closest competitor and the central ablation. Penalties trade objective quality against violation through a weight that must be tuned. Near the critical condition, small output errors cause full outage, so we expect a residual violation rate that cannot be tuned to zero without shrinking the front. If the penalty network reaches zero violations at equal hypervolume, the feasibility map is unnecessary. This one comparison decides the proposal.
Why not just sweep the fractional allocation exponent or a single rate knob? A one-parameter family can only trace a curve, while three objectives define a surface. Static coefficients also suffer the high-SNR outage floors that An Energy Efficient Design of Hybrid NOMA Based on Hybrid SIC with Power Adaptation documents, where failure probability plateaus at $10^{-1}$ or $10^{-2}$ without adaptation. The hypervolume gap against the swept heuristic measures how much is lost in practice.
Why not just apply max-min constrained multi-objective reinforcement learning? Power allocation within a time slot has no meaningful state transitions. That approach's expected-cost constraints permit per-slot violations, and it returns one fair point rather than a navigable front. We include it only as a single-point reference at the max-min preference.
Proposed Experiments
All experiments use the Rayleigh fading disc model of Ding et al., with users placed randomly in a disc and path loss applied, plus a multi-carrier extension with independent fading per subcarrier. We sweep three settings: $K \in \{2, 3, 4, 6\}$ users per cluster, 1 to 16 subcarriers, and normalized channel-estimation error variance from 0 to 0.1. The metrics are hypervolume and IGD+ (inverted generational distance plus, the average distance from the reference front to the learned one), both measured against the solver's reference front. We also report constraint-violation rate, empirical outage against target, distance between the requested preference direction and the realized trade-off (the mean-Euclidean-distance metric of the Hyper-Transformer paper), and inference latency.
There are three decisive checks. On two-user perfect-knowledge channels, the learned front must match the closed-form fronts of Xing et al. If it does not, the training is broken. In the multi-carrier, imperfect-knowledge setting, the feasibility-map network must beat the penalty ablation on violation rate at equal or better hypervolume. As estimation error grows, the calibrated backoff must hold empirical outage at its target, and the hypervolume loss relative to perfect knowledge must be reported. If the map wins only in the trivial single-carrier case, the claim fails.
References
- On the Performance of Non-Orthogonal Multiple Access in 5G Systems with Randomly Deployed Users
- Dealing With Structure Constraints in Evolutionary Pareto Set Learning
- A Survey on Evolutionary Constrained Multiobjective Optimization
- Immune Clonal Multi-Objective Optimization Algorithm for Constrained Optimization
- Learning the Pareto Front with Hypernetworks
- A Hyper-Transformer model for Controllable Pareto Front Learning with Split Feasibility Constraints
- Beam Design and User Scheduling for Non-Orthogonal Multiple Access with Multiple Antennas Based on Pareto-Optimality
- Optimal Throughput Fairness Tradeoffs for Downlink Non-Orthogonal Multiple Access Over Fading Channels
- minPIC: Optimal Power Allocation in Multi-User Interference Channels
- Constrained Multi-Objective Reinforcement Learning with Max-Min Criterion
- Energy-Efficient Resource Allocation in NOMA Heterogeneous Networks
- Constrained Multi-Objective Optimization With Deep Reinforcement Learning Assisted Operator Selection
- Deep Learning Based Radio Resource Management in NOMA Networks: User Association, Subchannel and Power Allocation
- An Energy Efficient Design of Hybrid NOMA Based on Hybrid SIC with Power Adaptation