A reconfigurable intelligent surface (RIS) is a flat panel of hundreds of cheap reflecting elements. Each element shifts the phase of an incoming radio wave, which lets one surface steer a beam toward a phone for communication and also illuminate objects for radar-style sensing. Real surfaces allow only a few discrete phase settings per element, typically 1 or 2 bits. Designers trade communication rate against sensing quality by searching over these settings. The sensing half of that trade is scored against a map of which locations hold targets. In a real room, warehouse, or street, that map is always incomplete. Some reflectors are surveyed. Many are not: parked cars, furniture, people. Current designs score any return from an unsurveyed location as a false alarm. The optimizer is therefore rewarded for steering energy away from real objects, and the rate-versus-sensing curve it reports is wrong in a systematic direction.
We propose to treat the sensing objective as a positive-unlabeled (PU) learning problem. In a PU problem, known positives are available but the rest is an unlabeled mixture of positives and true negatives, not a set of confirmed negatives. We then run sample-efficient multi-objective search directly over discrete phase configurations with this corrected objective. The claim is concrete. With naive negatives, the recovered Pareto front (the set of designs where neither rate nor sensing can improve without hurting the other) is biased. It is biased toward beams that avoid hidden scatterers, and the bias grows as the fraction of labeled scatterers falls. A PU-corrected objective should remove most of this bias at 1–2 bit resolution. Whether coarse quantization erases the small ranking margins that PU corrections depend on is an open question the file flags explicitly, and we answer it empirically.
Core Claim and What Is New
No paper in the file spans all three ingredients. Three wireless papers form the closest line of work, and all of them assume the scene is fully specified. Passive Beamforming Design for RIS-Enabled ISAC maps the rate-versus-detection trade-off for extended targets, but it uses continuous phases, alternating optimization, and binary hypothesis tests with known targets. MetaRadar jointly optimizes waveform and phase shifts for multi-target detection through a semidefinite relaxation. It assumes targets and non-targets are fixed in advance. Learning Beamforming Codebooks for Active Sensing learns discrete codebooks (VQ-C) but optimizes one scalar localization loss. These papers supply the physical setting and the discrete-control machinery. Our proposal adds two things none of them has. It uses a sensing objective that stays valid when most scatterers are unlabeled, and it searches for a front directly over discrete configurations.
The negative-sampling literature motivates the correction. LipoPU predicts lipid-binding protein pockets with a ranking-based PU objective and reports PR-AUC 0.839 versus 0.651 for a model that treats unlabeled pockets as negatives. PepBenchmark shows that naive negative selection inflates ROC-AUC through memorization. Both come from biology, so neither says anything about beam design. We borrow the lesson and the estimators, not the numbers. The specific estimator is Positive-unlabeled AUC Maximization under Covariate Shift. Covariate shift is intrinsic in our setting: each phase configuration changes which locations are illuminated, so the distribution of returns moves with the decision variable.
Research Program
The simulated scene contains $K$ true scatterers with extended physical cross-sections, following the extended-target model of the ISAC paper above. Only a fraction $c$ of them, the label frequency, is revealed to the designer. For a phase configuration $\theta \in \{0, 2\pi/2^b, \dots\}^N$, with $N$ elements at $b$ bits, a fixed matched-filter detector produces a score for each location cell. Sensing quality is the AUC of those scores separating target cells from empty cells.
The naive objective computes this AUC with labeled cells as positives and all other cells as negatives. The PU-corrected objective uses the fact that the unlabeled pool is a mixture with target prior $\pi$. The observed positive-versus-unlabeled AUC is related to the true AUC by
$$\mathrm{AUC}_{PN}(\theta) = \frac{\mathrm{AUC}_{PU}(\theta) - \pi/2}{1-\pi},$$
where $\mathrm{AUC}_{PU}$ is the AUC of labeled positives against unlabeled cells. The relation comes from decomposing the unlabeled cells into a $\pi$ fraction of positives, which contribute AUC $1/2$ against labeled positives, and a $1-\pi$ fraction of negatives. It assumes labeled targets are a random sample of all targets. Under covariate shift that assumption fails, and we use the importance-weighted variant from the covariate-shift PU paper instead. We estimate $\pi$ from one small fully surveyed patch. We treat that estimate as the fragile part and stress-test it.
The second objective is the communication rate to one or more users. A third, optional objective is the number of distinct configurations used per sensing frame, which stands in for the control-link cost that VQ-C highlights.
The search runs directly over discrete configurations. To keep dimension manageable we tie elements into tiles, for example 64 tiles at 2 bits, giving a 128-bit decision vector. We then run multi-objective Bayesian optimization with tree-ensemble surrogates, following Thebelt et al.. That work reports Pareto fronts found more efficiently than NSGA-II on constrained problems, and tree ensembles handle categorical inputs natively. As a larger-scale variant, we use regionalized local search from Daulton et al. with element-level control. The method under test is PU-corrected tree-ensemble multi-objective search.
The analysis asks two questions. First, how does front bias scale with $c$ and with bit depth $b \in \{1,2,3\}$? Front bias is the gap between the front the designer believes it has and the front evaluated against the full scene. Second, does the PU benefit survive 1-bit control? The file's spatial-averaging results, the $N^2$ SNR scaling under 1-bit control (Jung et al.) and the roughly 1.6× aperture compensation for 1-bit secrecy (Shi et al.), suggest a testable prediction. If ranking margins are also averaged out, then a 1-bit surface scaled up by a similar factor should recover the 2-bit PU gain.
Baselines
The naive reference point is the same tree-ensemble multi-objective search with the naive objective, where unlabeled cells count as negatives. This is the ablation that isolates the PU correction. Optimizer, evaluation budget, discrete space, and detector are identical to the proposed method, and only the sensing score differs. The proposal beats it if hypervolume under the ground-truth scene is higher, and front bias lower, by more than seed-to-seed variation at $c \le 0.5$. At $c = 1$ the two methods must coincide, which serves as a correctness check.
The strongest existing method for the physical trade-off is alternating optimization with successive convex approximation from the ISAC passive beamforming paper. It is run with continuous phases, rounded to $b$ bits, and swept over a weight grid to trace a front. That paper reports maintaining 90% detection probability for extended targets, against as low as $10^{-5}$ for point-target designs. That result comes from its own fully labeled simulator. No reported result was found on Lacuna for any discrete-phase ISAC front under partial labels, so we reimplement it inside our simulator. It receives the same channel model and the same number of objective evaluations, counting each weight's run. We also report it with the PU-corrected score inside its objective, so that the optimizer and the objective can be credited separately.
The learned discrete baseline is VQ-C from the active sensing codebook paper. It reports needing only 14 bits per control signal at codebook size 10,000 while matching codebook-free sensing, but it gives no localization error in meters. We adapt it to our scene by training its codebook with a scalarized rate-plus-sensing loss and sweeping the weight. NSGA-II over the same tile encoding, at an equal evaluation budget, is the evolutionary baseline.
An oracle run with all scatterers labeled gives the upper bound. Every method gets the same number of simulator calls and the same hyperparameter tuning budget on a separate set of scenes, and all methods are evaluated on the same held-out scenes over 20 seeds.
Why not just…?
Why not just optimize continuous phases and round? Rounding is cheap and sometimes good enough. The saturation results by Huang et al. and Di et al. show that 2–3 bits capture most energy efficiency and sum rate. Those results concern bulk SNR and rate, not detection of weak extended targets, which face double path loss through the surface. The file also reports that VQ-C beat continuous-to-discrete heuristics. We do not assume the outcome. The rounded baseline is run at every $b$. If it matches direct discrete search at 2 and 3 bits, we will say that discrete search matters only at 1 bit.
Why not just use weighted sums? Hayes et al. argue that scalarization cannot reach non-convex parts of a front. Gharib et al. report roughly an order of magnitude more evaluations for scalarized search to reach comparable fronts in path-following control. Whether our rate-versus-sensing front is actually non-convex is unknown. The scalarized VQ-C and alternating-optimization sweeps measure this directly: if they reach matched hypervolume at matched budget, the multi-objective machinery is not needed here.
Why not just ignore unlabeled cells and only reward energy on known targets? This is the most serious alternative. A positive-only objective avoids label contamination entirely. Its cost is that it gives no false-alarm control and no incentive to illuminate unsurveyed space, so it should find beams that concentrate on known targets and miss hidden ones. We include it as a baseline and measure the hidden-target detection rate separately from the overall AUC. If it matches the PU method on both, the PU machinery adds nothing and the paper's conclusion becomes "drop the negatives."
Why not just add elements or transmit power to drown out the label noise? Label contamination is a bias in the objective, not noise, so more SNR should sharpen the optimizer's avoidance of hidden scatterers rather than fix it. The sweep over $N$ from 64 to 1024 tests this. The prediction is that naive front bias persists or grows with $N$. If it vanishes, the problem is not worth solving with PU methods.
Proposed Experiments
The simulator follows the RIS-aided ISAC geometry of the passive beamforming paper, with Rician fading, extended targets, and a 30 m × 70 m area as in the LSTM active sensing study. We sweep label frequency $c \in \{0.1, 0.25, 0.5, 1\}$, bits $b \in \{1,2,3\}$, and elements $N$ from 64 to 1024. Metrics are hypervolume under the ground-truth scene, front bias, hidden-target detection rate at fixed false-alarm rate, and objective evaluations needed to reach 95% of oracle hypervolume.
Three checks decide the proposal. First, at $c=0.25$ and $b=2$, PU-corrected search must beat the naive-negative ablation on true hypervolume. Otherwise the core claim fails. Second, varying the $\pi$ estimate by ±50% must not flip that ordering. Otherwise the method is too fragile for deployment. Third, the 1-bit experiment settles the open quantization tension: either PU gains collapse at 1 bit regardless of $N$, or they return once the aperture is scaled. Either outcome is reportable. A secondary transfer check applies the same corrected-objective search to the multi-task over-the-air learning setting of Quantized Analog Beamforming Enabled Multi-task FL, using MNIST and CIFAR-10 clients with one-class, positive-only labels. That paper reports no accuracy numbers, so we will state "no reported result found on Lacuna" and compare only against our own reimplementation.
References
- Passive Beamforming Design for Reconfigurable Intelligent Surface Enabled Integrated Sensing and Communication
- MetaRadar: Multi-Target Detection for Reconfigurable Intelligent Surface Aided Radar Systems
- Learning Beamforming Codebooks for Active Sensing With Reconfigurable Intelligent Surface
- LipoPU: Pocket-level Prediction of Lipid-Protein Interactions via Positive-Unlabeled Learning
- PepBenchmark: A Standardized Benchmark for Peptide Machine Learning
- Positive-unlabeled AUC Maximization under Covariate Shift
- Multi-Objective Constrained Optimization for Energy Applications via Tree Ensembles
- Multi-Objective Bayesian Optimization over High-Dimensional Search Spaces
- On the Optimality of Reconfigurable Intelligent Surfaces (RISs): Passive Beamforming, Modulation, and Resource Allocation
- Secure Outage Analysis of RIS-Assisted Communications With Discrete Phase Control
- Energy Efficient Multi-User MISO Communication Using Low Resolution Large Intelligent Surfaces
- Hybrid Beamforming for Reconfigurable Intelligent Surface based Multi-User Communications: Achievable Rates With Limited Discrete Phase Shifts
- A practical guide to multi-objective reinforcement learning and planning
- Multi-Objective Optimization of a Path-following MPC for Vehicle Guidance: A Bayesian Optimization Approach
- Active Sensing for Localization with Reconfigurable Intelligent Surface
- Quantized Analog Beamforming Enabled Multi-task Federated Learning Over-the-air