Information theory has a counterintuitive result about interference. If a neighbouring transmitter's signal arrives much stronger than your own, your receiver can decode that neighbour's message first, subtract it, and then decode its own signal as if the neighbour were absent. For symmetric networks this "very strong interference regime" even has a clean entry condition that does not depend on how many users there are. The catch is that the condition holds only for idealized channels: every cross link equally strong, perfect phase alignment, and ideal codes. Real channels are asymmetric and slightly out of sync, and they are decoded with finite-length codes. For them, nobody knows which transmit powers, code scalings and decoding orders actually reach interference-free rates. The only reliable way to find out is expensive link-level simulation, whose run time depends heavily on which decoding strategy is being tested.
We propose treating this search as a parallel black-box optimization problem, with one specific change. Instead of modelling measured throughput directly, the surrogate (the statistical model that predicts outcomes of untried configurations) models the smooth quantities inside each decoding strategy: the effective signal-to-noise margin at every decode-and-subtract step. It then recombines them through the known max-over-strategies, min-over-steps structure that information theory supplies. Throughput jumps abruptly when a decoding step starts or stops succeeding, and generic parallel Bayesian optimization handles such cliffs badly. We claim that the regime structure should live inside the surrogate rather than be approximated by distance-based batch diversity. Doing so lets many simulations run concurrently and asynchronously without wasting them on the cliff edges.
Core Claim
Let $x$ be a network configuration: per-user transmit powers, lattice scaling factors, and a decoding strategy $s$ at each receiver. Examples of $s$ are treating interference as noise, successively peeling interferers, or decoding their aligned sum. The simulator's throughput $f(x)$ is discontinuous. The per-step margins $m_{s,j}(x)$, meaning the effective SNR seen at step $j$ of strategy $s$, are smooth in the continuous parameters. The achievable rate therefore has the form
$$ f(x) = \max_{s} \; \min_{j} \; r\big(m_{s,j}(x)\big), $$
where $r(\cdot)$ is a known, monotone map from margin to supportable rate, calibrated once per code family. We place Gaussian process models on the $m_{s,j}$, not on $f$. We draw posterior samples of every margin and push them through this formula to get Thompson samples of $f$. Each idle worker is sent the maximizer of a fresh sample as soon as it frees up.
A single transmit configuration can be decoded under several strategies. Each simulation therefore returns margins for several $s$ at once, and the acquisition decides both what to transmit and which decoders are worth their compute. The prediction is concrete. At equal wall-clock budget, this regime-aware loop should approach the interference-free point-to-point rate on asymmetric, phase-perturbed $K$-user channels faster than parallel Bayesian optimization on $f$ itself. The advantage should grow as the batch size and the number of users grow.
Grounding in Prior Work
The interference side supplies the outer structure and our sanity targets. Capacity of Symmetric K-User Gaussian Very Strong Interference Channels shows that layered lattice codes let a receiver decode the sum of interferers as one object. With unit direct gains and cross gain $a$, every user reaches $\tfrac12\log(1+P)$ once $a^2 \ge (P+1)^2/P$. For $P=10$, $K=10$ that is $a^2 \ge 12.1$, versus roughly $2\times10^8$ for naive one-by-one decoding. The same paper flags asymmetric channels and phase or timing offsets as open problems. That paper motivates our work and bounds it: we do not attempt a new capacity theorem for the asymmetric case, only an empirical search for good operating points where no theorem exists.
Binary Fading Interference Channel With No CSIT gives exact sum capacity in its weak regime, $2p_d(1-p_c)$, and its strong regime, $p_d+p_c-p_dp_c$. It is a cheap testbed where the regime boundaries and the optimum are known exactly. On the Spectral Efficiency of Noncooperative Uplink Massive MIMO Systems shows with random-matrix analysis that choosing between simultaneous decoding and treating interference as noise beats either one alone. It reports gains of 2.2 and 0.45 bps/Hz over matched and MMSE filtering. This is evidence that the decoding-strategy choice matters, but the analysis relies on closed forms, which we explicitly do not assume.
The optimization side supplies the machinery we reuse. Parallelised Bayesian Optimisation via Thompson Sampling proves that asynchronous Thompson sampling across $M$ workers matches sequential regret rates, which removes the synchronization barrier; we adopt its dispatch rule unchanged. Maximizing acquisition functions for Bayesian optimization shows that greedy batch construction of Monte Carlo acquisition functions carries a $(1-1/e)$ guarantee and beats joint optimization over all batch points. We use it as the synchronous comparison.
Two methods enforce batch diversity through distance in configuration space. Batch Bayesian Optimization via Local Penalization calibrates an exclusion radius with a global Lipschitz constant, a bound on how fast the function can change. Batched Gaussian Process Bandit Optimization via Determinantal Point Processes samples diverse batches from the posterior covariance. Both assume a roughly stationary objective, whose smoothness is the same everywhere. A throughput surface with capacity cliffs violates that assumption, and our proposal differs precisely in not relying on it.
The closest applied work is Analysis of a decentralized Bayesian optimization algorithm for improving spatial reuse in dense WLANs (INSPIRE). It tunes Wi-Fi transmit power and detection thresholds with per-access-point Gaussian processes, reporting a 400% throughput gain over defaults in ns-3. It treats interference purely as something to back away from, runs in synchronous rounds, and truncates memory to 50 points. We differ on all three counts: cancellation-based strategies, an asynchronous parallel loop, and a structured surrogate.
Research Program
We will build a link-level simulator for asymmetric $K$-user Gaussian interference channels on GPU-accelerated Sionna. It will implement three receiver strategies:
- Treating interference as noise.
- Successive decode-and-subtract.
- Aligned lattice sum decoding.
Cross gains will come from path-loss-plus-shadowing draws. We will add controllable residual phase offsets, carrier frequency offsets, and finite blocklength. The simulator will expose per-step post-equalization margins alongside block error rate.
The run time of an evaluation depends on the strategy. Successive peeling costs up to $K-1$ decodes, while aligned sum decoding costs one. That variation is exactly the straggler pattern that makes synchronous batching wasteful.
On top of the simulator we will implement three pieces:
- The composite surrogate. It uses independent or lightly coupled Gaussian processes per margin, with $r(\cdot)$ fitted from single-link error curves.
- Asynchronous composite Thompson sampling. Each dispatch jointly selects a transmit configuration and the subset of decoders to run, weighted by predicted improvement per unit of decode time.
- A diagnostic of batch waste. It measures the fraction of evaluations whose outcome lies within one margin-noise standard deviation of a decoding-success boundary but on the losing side.
The analysis question is whether the composite model's calibration near regime boundaries is what drives any speedup. We check this by comparing posterior coverage of $f$ near boundaries under the composite and the flat models.
Baselines
The naive reference points involve no search over cancellation at all. The first is treating interference as noise with per-user power control, tuned by sequential Bayesian optimization on the same simulator. The second is time division among the $K$ users, which caps each user's rate at a $1/K$ share. On the Spectral Efficiency of Noncooperative Uplink Massive MIMO Systems shows both are dominated by a strategy-switching scheme when interference is strong. We run them anyway, because on asymmetric channels with phase noise that dominance is exactly what is in question.
The strongest analytic competitor is an "analytic plug-in." It takes the symmetric threshold from Capacity of Symmetric K-User Gaussian Very Strong Interference Channels and applies it to the weakest cross link. It sets powers and scalings so that link clears $(P+1)^2/P$, then evaluates the result once in the simulator. It costs almost nothing. No reported result was found on Lacuna for this plug-in on asymmetric or phase-perturbed channels. The only reported reference is the idealized threshold itself, $a^2\ge12.1$ at $P=10$, $K=10$.
The strongest black-box competitor is asynchronous Thompson sampling on the flat throughput $f$, as in Parallelised Bayesian Optimisation via Thompson Sampling. Its authors report only regret curves with no scalar values. On synchronous batch methods and this wireless task, no reported result was found on Lacuna. We also run greedy $q$-UCB from Maximizing acquisition functions for Bayesian optimization, local penalization from Batch Bayesian Optimization via Local Penalization, and DPP sampling from Batched Gaussian Process Bandit Optimization via Determinantal Point Processes. An INSPIRE-style windowed Gaussian process (50 points) tests whether memory truncation hurts near boundaries.
All methods share the same simulator seeds, the same worker count $M \in \{4, 16, 64\}$, and the same wall-clock budget. They also share the same kernel family and hyperparameter-fitting routine, and each gets an equal budget of 20 random restarts for acquisition optimization.
Two ablations isolate the key ingredient. The first keeps asynchronous dispatch but uses a flat surrogate that is simply given the strategy label as an extra input. The second keeps the composite surrogate but runs synchronous greedy batches. To beat the strongest competitor, the full method must reach 95% of the best rate found by any method in at most half the wall-clock time of flat asynchronous Thompson sampling, across at least three of four channel-asymmetry levels.
Why not just…?
Why not just apply the analytic threshold to the worst link and skip optimization? This is the most serious threat, and it may win on nearly symmetric channels with small phase error. The file offers no evidence either way for asymmetric channels. The symmetric theorem's own authors list asymmetry and offsets as unresolved, which suggests the plug-in will either over-provision power or miss the regime. The experiment settles this directly. We sweep the asymmetry spread and phase-offset level, and report where the plug-in's gap to the best-found configuration exceeds 0.1 bit per user. If it never does, the honest conclusion is that optimization is unnecessary in that range.
Why not just run a good generic parallel method, such as asynchronous Thompson sampling, on throughput directly? Posterior sampling already diversifies batches, so structure may add little. The worry is specific to this problem. The throughput surface has cliffs where a decoding step starts failing. A stationary Gaussian process on $f$ blurs each cliff into a slope, and parallel workers then land on the losing side of it. The waste diagnostic and the flat-surrogate-with-labels ablation measure exactly this. If flat asynchronous Thompson sampling matches the composite method at equal wall-clock time, the core claim is falsified.
Why not just treat interference as noise and tune powers, or divide time among users? Both are simple and robust. However, On the Spectral Efficiency of Noncooperative Uplink Massive MIMO Systems and Binary Fading Interference Channel With No CSIT show both are strictly suboptimal under strong interference. Time division also pays a $1/K$ rate penalty, whereas the symmetric theorem shows cancellation removes that penalty entirely. They remain in the study as floors, and the gap to them tells us how much cancellation is worth under realistic impairments.
Why not just use local penalization with a locally estimated Lipschitz constant? That would patch the non-stationarity through the penalty radius rather than through the model. It still cannot tell which side of a boundary a point lies on, only how steep the surface is nearby. We include a per-region Lipschitz variant of local penalization so the comparison is not against a straw man.
Proposed Experiments
The first experiment is a recovery check on the symmetric $K$-user Gaussian channel with no impairments. Here the answer is known: full rate $\tfrac12\log(1+P)$ beyond $a^2\ge(P+1)^2/P$, which is 12.1 at $P=10$. Every method should find it, and we record the evaluations needed as $K$ goes from 3 to 10. If the composite method is not faster here, it will not be faster elsewhere.
The second experiment uses the binary fading interference channel. Exact sum capacity is known in its weak and strong regimes, so simple regret can be measured against ground truth. Sweeping $p_c$ across the moderate band tests behaviour right at regime boundaries.
The main experiment is the asymmetric, phase-perturbed $K$-user link-level simulation. We report four metrics: the fraction of isolated point-to-point rate achieved versus wall-clock time, worker idle fraction, evaluations wasted on the failing side of boundaries, and the final configurations reached. Configurations are reported as how many receivers ended in each decoding strategy.
A small synthetic control tests whether the benefit comes from the known max-min structure rather than from wireless specifics. It builds objectives as a max over several smooth Hartmann-6 and Branin pieces with min-constraints.
Two outcomes are decisive. The first is whether the waste diagnostic falls for the composite surrogate while falling much less for the flat-with-labels ablation. That would confirm the structured model, not the extra information, as the cause. The second is whether the composite method's advantage over flat asynchronous Thompson sampling grows with worker count $M$. If it does not, the result says only that the composite model is a slightly better surrogate, not that it relieves the parallel bottleneck.
References
- Capacity of Symmetric K-User Gaussian Very Strong Interference Channels
- Binary Fading Interference Channel With No CSIT
- On the Spectral Efficiency of Noncooperative Uplink Massive MIMO Systems
- Parallelised Bayesian Optimisation via Thompson Sampling
- Maximizing acquisition functions for Bayesian optimization
- Batch Bayesian Optimization via Local Penalization
- Batched Gaussian Process Bandit Optimization via Determinantal Point Processes
- Analysis of a decentralized Bayesian optimization algorithm for improving spatial reuse in dense WLANs
- PyJama: Differentiable Jamming and Anti-Jamming with NVIDIA Sionna