From Confounding to Counterfactuals: A Cross-Layer Causal Model for RAN Scheduling

From Confounding to Counterfactuals: A Cross-Layer Causal Model for RAN Scheduling Artin Elhamirad∗ , Christo Kurisummoottil Thomas∗ , Ali Tajer† ∗ Department of Electrical & Computer Engineering, Worcester Polytechnic Institute, Worcester, MA, USA † Department of Electrical, Computer, and Systems Engineering,

arXiv:2611.00635v1 [cs.NI] 31 Oct 2026

Rensselaer Polytechnic Institute, Troy, NY, USA Email: aelhamirad@wpi.edu, cthomas2@wpi.edu, tajera@rpi.edu Abstract—Data-driven control in radio access networks (RANs) selects resource allocations using associations between control actions and key performance indicators (KPIs) learned from network telemetry. Because the shared radio environment influences both the scheduler’s decisions and the resulting performance, a bias known as confounding, such associations overstate the effect of an allocation and misdirect resources toward user equipment (UEs) that cannot benefit. Moreover, individual UEs respond heterogeneously to the same decision, so even a corrected average effect does not identify which UEs gain from additional resources. While causal inference has recently been applied to RANs, existing approaches do not model the cross-layer mechanisms that connect an allocation to its outcome. In this paper, a hierarchical structural causal model (H-SCM) spanning the PHY, MAC, and network layers is proposed that separates cell-load confounding from UEspecific response heterogeneity. In particular, a counterfactual scheduling problem is formulated in which physical resource block (PRB) allocations are selected using per-UE counterfactual goodput rather than observational conditional averages. This problem is solved via a pipeline combining constrained causal discovery, structural-equation fitting, and abduction–action– prediction. Simulations demonstrate that naive observational estimation overstates the full-PRB response by 73.3%, whereas adjustment for cell-load confounding reduces the relative error to 9.8%. Held-out interventions are predicted with a coefficient of determination of 0.988. Counterfactual scheduling achieves a mean relative loss in total cell goodput below 0.5% against the ground-truth optimum, compared with 2.0–2.9% for a neuralnetwork baseline.

I. I NTRODUCTION Network control in modern cellular radio access networks (RANs) relies on telemetry and data-driven artificial intelligence (AI) models to support resource-allocation decisions such as physical resource block (PRB) allocation among network slices and scheduling-policy selection [1]. These models learn statistical associations between network states, control actions, and key performance indicators (KPIs), and use these to select the actions that maximize performance objectives. However, such statistical associations can be misleading when the radio environment influences both the scheduler’s decisions and the resulting KPIs. For example, low cell load can simultaneously increase the PRBs available to a user equipment (UE) and improve its link quality through reduced co-channel interference. Consequently, the higher goodput observed at higher PRB allocations may partly reflect lower cell load rather than the effect of allocating more PRBs. This situation, in which a shared environmental variable influences both the control action and the outcome, is known as confounding [2]. A controller that ignores it can misattribute environmental effects to the action being optimized.

While this confounding can be corrected by accounting for the congestion state when estimating the allocation effect, such a correction does not by itself enable UE-specific control. The corrected estimate characterizes only the populationaverage response, whereas UEs differ in propagation, device, and traffic conditions. The same increase in PRB allocation may therefore yield a negligible gain for a channel-limited UE and a substantial gain for a PRB-limited UE. Per-UE resource allocation instead requires predicting the outcome that a specific UE would attain under an alternative allocation, given the conditions it actually experienced. Such a unitspecific prediction is known as a counterfactual [2]. This work therefore addresses the following question: How can a RAN estimate the causal effects of resource-allocation decisions from observational telemetry, free of confounding by the network state, and predict the counterfactual response of an individual UE to an alternative allocation? A. Related Work Traditional resource-allocation approaches optimize weighted-throughput and proportional-fair objectives [3]. More recently, there has been growing interest in datadriven control for scheduling, resource allocation, and slice management through the RAN Intelligent Controller (RIC), including machine-learning-based xApps and reinforcementlearning methods [1] and [4]. Both classes of formulations optimize network-performance objectives without explicitly distinguishing statistical association from the interventional effect of a scheduling decision. A principled approach to overcoming these limitations is causal inference, in which structural causal models (SCMs) formalize the distinction between observational, interventional, and counterfactual queries [2], and causal structure can be learned from data [5]. In this framework, confounding biases the estimated effect of an allocation, whereas effect modification means that this effect differs across UEs or network states, and a variable such as congestion can play both roles [6]. In wireless networks, causal reasoning has been advocated as a foundation for AI-native systems, with applications including resilient beamforming, digital-twin modeling, and intent management [7]. CausalTune adjusts for environmental and operational confounders in RAN configuration tuning [8], while BLINC uses a large language model (LLM) to supply domain-knowledge priors for Bayesian structure learning in context-specific configuration [9]. At the counterfactual level,

conformal methods have been used to estimate the KPI modeled as a function of γeff and j ⋆ . We select the highestthat would have resulted from deploying a different RAN rate MCS satisfying the BLER target from the 3GPP MCS application [10], and the interaction between a user, an LLM table [14], which specifies the corresponding modulation order agent, and the network has been modeled as an SCM to evaluate bmod and target code rate R, which determine the nominal PHY alternative formulations of a user intent [11]. These methods spectral efficiency, ηPHY = bmod R, measured in bits/s/Hz. reason at the level of configurations, applications, or intents [8]– In the MAC layer (L2), we model the number of PRBs [11], without modeling the cross-layer mechanisms that connect allocated to the UE by a load-dependent scheduling rule, a resource-allocation decision to its outcome. Moreover, each T = min{NPRB , max{1, ⌊NPRB (1 − αL L)⌋ + εT }} , (2) addresses either confounding or counterfactual evaluation in where NPRB is the total number of available PRBs, αL ∈ (0, 1) isolation, whereas UE-specific allocation requires both. controls the scheduler’s sensitivity to load, and εT captures B. Contributions slot-level scheduling variability. Given the allocated PRBs, the The main contribution of this paper is a cross-layer causal MAC-layer service rate is further shaped by retransmission framework for RAN resource allocation that distinguishes celloverhead and device-specific radio-frequency (RF) variation. load confounding from UE-specific response heterogeneity. We We approximate the overhead of HARQ retransmission proceformulate a counterfactual scheduling problem that allocates dures [15] using a truncated expected-retransmission surrogate. PRBs according to each UE’s predicted goodput gain and Let KH be the maximum number of retransmissions considered develop a hierarchical structural causal model (H-SCM) in the model, and Nretx ∈ {0, . . . , KH } the number of including the physical (PHY), medium access control (MAC), HARQ retransmissions. Since Nretx is a nonnegative integerand network layers to predict these individual responses. Simvalued random variable, its expectation equals the sum of ulations demonstrate that accounting for cell load reduces the its tail probabilities. Under the simplifying assumption that relative error in estimated goodput under full-PRB allocation every transmission attempt fails independently with the same from 73.3% to 9.8%. Allocating PRBs based on each UE’s probability p, at least k retransmissions are required exactly predicted response achieves a mean relative loss in total cell when the first k attempts fail, so Pr(Nretx ≥ k) = pk and goodput below 0.5% against the ground-truth optimum. KH KH X X E[N ] = Pr(N ≥ k) = pk . (3) retx retx II. S YSTEM M ODEL AND S CHEDULING M OTIVATION k=1 k=1 Consider a single-cell downlink RAN serving a set Us of This expression approximates the retransmission overhead K UEs in each scheduling interval. We model the system without explicitly modeling HARQ combining. across the PHY, MAC, and network layers, whose variables Device-specific RF variation, arising from differences in and mechanisms, such as link adaptation, hybrid automatic receiver hardware implementation such as front-end noise repeat request (HARQ) retransmission, and queueing, jointly figure and component tolerances across UE models, is repdetermine the application-layer goodput. Cross-layer dependen- resented by ξ 2 hw ∼ N (0, σhw ). The MAC service rate is cies are modeled on a per-UE basis, while cell-level quantities µ = C T η 0 PHY max{1+κhw ξhw , 0.05}/(1+κHARQ E[Nretx ]), such as the congestion state and the available PRB budget are where C converts to Mbps and the κ coefficients scale 0 shared among the UEs scheduled in the same interval. Let hardware variation and HARQ overhead. Although hardware L ∈ [0, 1] represent a cell-level congestion state summarizing impairments physically degrade the received SINR [3], we coupled serving-cell resource pressure and co-channel activity approximate their net effect through an assumed MAC-layer in the surrounding radio environment. As L increases, the radio multiplier, which keeps the device-specific pathway separate resources available for serving-cell scheduling decrease, and from the propagation pathway through ξ . sf the co-channel interference experienced by the UE grows. In the network layer (L3), traffic burstiness is characterized 2 A. Wireless RAN System Model by ξtb ≜ Ca2 = σA /µ2A , the squared coefficient of variation 2 For notational simplicity, we omit the UE index u in this of the UE’s inter-arrival times A, where µA and σA are section. In the PHY layer (L1), let γ0 be the nominal received the mean and variance of A. This is a standard variability signal-to-noise ratio (SNR), fD the Doppler spread, and τrms measure in queueing theory [16]. Poisson arrivals have ξtb = 1, the delay spread. The UE-specific shadow-fading offset is while ξtb > 1 indicates greater inter-arrival variability. Using 2 modeled as ξsf ∼ N (0, σsf ) [12]. The effective signal-to- the GI/G/1 heavy-traffic approximation for mean waiting interference-plus-noise ratio (SINR) can be modeled as time [16], assuming no service-time variability conditional on γeff = γ0 + ξsf − βL L − ∆D (fD ) − ∆τ (τrms ) + εγ . (1) the modeled MAC rate, and applying Little’s law, we model the mean queue backlog as  Here, we model the penalties as ∆D (fD ) = aD log2 (1+fD /f0 ) ρ2 ξtb and ∆τ (τrms ) = aτ log2 (1 + τrms /τ0 ), where aD , aτ > 0 Q= + ε3 , (4) 2(1 − ρ) + are penalty coefficients in dB and f , τ > 0 are reference 0

0

scales. The parameter βL > 0 scales the interference contribu- where ρ = min{λ/µ, ρmax }, λ is the rate of the queued traffic tion of the load state, consistent with cellular load-coupling component that loads the UE’s buffer and sets its backlog models [13], and εγ is a residual noise term accounting for through ρ, ρmax < 1 caps utilization, ε3 captures residual unmodeled effects on γeff . The block error rate (BLER) p is queueing variation, and [x]+ = max{x, 0}. Further, we model

the packet-loss probability, which captures residual decoding errors and queue buildup, as  
[Q − qth ]+ 1+KH Ploss = 1 − 1 − p exp − . (5) qscale In (5), the first factor is the probability that decoding succeeds within KH + 1 attempts under the same independent-attempt assumption as (3), and the second factor models the probability of avoiding queue-induced drops as decreasing smoothly once the backlog Q exceeds a threshold qth . The parameter qscale controls the rate of this decrease. The two loss mechanisms are treated as independent. Finally, the application-layer goodput is Y = µ(1 − Ploss ), the goodput attainable by a saturated (fullbuffer) flow, i.e., the service capacity net of residual decoding and buffer-overflow losses. B. Conventional Scheduling Using the per-UE model, we now consider scheduling across the UEs Us , where T (u) ∈ T and Y (u) denote the PRB allocation and resulting goodput of UE u, respectively. The network state available to the scheduler when selecting the (u) (u) (u) PRB allocation is x(u) = [L, γ0 , fD , τrms ]T . A conventional scheduler assigns each UE a scheduling weight w(u) , which encodes its priority under the operator’s fairness policy, and allocates PRBs from a total budget B accordingly. A datadriven weighted sum-rate scheduler can then be formulated as h i X [3] max w(u) E Y (u) | T (u) = t(u) , x(u)

Fig. 1. Causal DAG spanning the PHY, MAC, and network layers. The DAG represents the simulated setting; operational RAN datasets may require additional nodes and dependencies.

The structural equations induce the directed acyclic graph (DAG) in Fig. 1. Acyclicity follows from the layer partition together with the feedforward processing order within each layer. The resulting observational distribution over context z and endogenous-variable values v = (vi )Vi ∈V , including PRB allocation and theQresulting cross-layer KPIs, factorizes as PM (z, v) = P (z) i P (vi | Pa(Vi )) , where each conditional P (vi | Pa(Vi )) is induced by the mechanism fi with Ui {t(u) } u∈Us (6) marginalized out. For deterministic mechanisms it is a point X s.t. t(u) ≤ B, t(u) ∈ T . mass. u In the H-SCM, cell load L is a common cause of the PRB In (6), the conditional expectation is estimated from historical allocation T and effective SINR γeff , creating the backdoor (u) telemetry. Since x includes cell load L, conditioning on this path T ← L → γeff → · · · → Y . Accounting for L blocks this state blocks the modeled confounding path. However, the tight path when estimating the population effect of PRB allocation. coupling between load and allocation limits observational supHerein, we define the do-operator do(T = t) that severs L → T port for some candidate allocations. The conditional mean also but leaves L → γ active, since forcing the scheduler does eff averages over unobserved UE-specific conditions. We propose not alter the UE’s interference environment. Cell load L can an H-SCM to infer these conditions from factual telemetry and also modify the goodput gain from additional PRBs. In our predict each UE’s response to alternative allocations. DAG, the UE-specific factors ξsf , ξhw , and ξtb modify this gain III. P ROPOSED H-SCM F RAMEWORK FOR C AUSAL without confounding the allocation effect and are preserved R ESOURCE A LLOCATION during unit-level counterfactual inference. We model the cross-layer cause-effect relations as an H-SCM, A. Three Rungs of Causal Inference defined as M = (U , V, F , PU ) [2], where each endogenous The H-SCM supports observational, interventional, and variable Vi ∈ V is determined by the structural equation Vi = counterfactual queries about PRB allocation and goodput [2]. fi (Pa(Vi ), Ui ) , where Pa(Vi ) represents its causal parents and At the observational rung, the mean goodput among UEs Ui ∈ U is an exogenous noise term. The model is hierarchical observed with allocation T = t is because its endogenous variables, comprising the PHY, MAC, µobs (t) = E[Y | T = t], (7) and network quantities from the effective SINR through the which, unlike the conditional mean in (6), does not condition allocation and service rate to the goodput, are partitioned by on L and is therefore confounded under the assumed DAG. the three protocol layers, V = V (1) ∪ V (2) ∪ V (3) , with no At the interventional rung, the backdoor structure established edges directed from higher to lower layers. The inputs to previously identifies the population response to setting T = t these mechanisms divide into context variables, which remain by adjustment for L as fixed within a scheduling epoch, and mechanism-level noise, µint (t) = E[Y | do(T = t)] = EL [E[Y | T = t, L]] . (8) which is redrawn in each realization. Let z collect the context variables, namely the congestion state L, the radio-link variables Computing µint (t) from observational data via (8) additionally {γ0 , fD , τrms }, and the unit-specific factors {ξsf , ξhw , ξtb }. The requires the overlap condition P (T = t | L) > 0 over the set U contains the noise terms {εγ , εT , ε3 }, and the remaining relevant load distribution [2]. Because the scheduler couples T mechanisms are deterministic given their parents. tightly to L, extreme allocations are rarely observed under

unfavorable loads, so the controlled interventions in our evaluation provide direct coverage of these allocations. At the counterfactual rung, the predicted goodput of UE (u) u under allocation t is Ŷt = Yt (û(u) ), obtained by fixing the allocation to T = t through do(T = t) and evaluating the fitted mechanisms with the UE-specific state û(u) , inferred from factual telemetry and held fixed. B. Counterfactual Scheduling Problem For each UE u, we select an allocation t(u) to maximize the weighted sum of counterfactual goodput under a shared PRB budget. The counterfactual scheduling problem is then X (u) (u) max w Ŷt(u)

TABLE I K EY S IMULATION PARAMETERS Parameter

Value

OFDM grid 76 subcarriers, 15 kHz spacing, 14 symbols Coding and MCS set LDPC with 512 information bits; 12 MCSs BLER target / HARQ limit 0.1 / 3 retransmissions PRB and load parameters NPRB = 100, αL = 0.58, βL = 18 dB Queue loading rate λ / utiliza- 5.5 Mbps / ρmax = 0.95 tion cap Queue-loss parameters qth = 5, qscale = 10

The two arms share identical exogenous draws. A candidate edge is retained if an effect remains significant after mediator blocking, but unblocked indirect paths can prevent interpreting {t(u) } it as a direct edge. Mediator selection uses the candidate graph u∈Us (9) and protocol order, without consulting the ground-truth edges. X s.t. t(u) ≤ B, t(u) ∈ T . 2) Structural Equation Fitting: For each endogenous variu able Vi , its structural mechanism is estimated from its causal A conditional causal response can already support UE-specific parents as (u) scheduling through the observed state x , but averages over N  2 1 X  (n) unobserved conditions. The objective in (9) additionally uses (n) fˆi = arg min vi − f Pa(vi ) . (11) each UE’s residual state inferred from factual telemetry to f N n=1 predict its response to alternative allocations, assuming that We fit each structural mechanism with a gradient boosting machine (GBM) using regression-tree base learners [18], which state persists. capture nonlinear dependencies and interactions among its C. Proposed Solution via a Causal Inference Pipeline We solve (9) in three stages to predict each UE’s goodput tabular parent variables. Telemetry fits use observable parents under alternative allocations while preserving its inferred in the recovered graph. Cross-fitting [19] computes each UE’s conditions. Accordingly, we recover the causal graph, fit the residual using a model trained without that UE. structural mechanisms, and infer each UE’s residual state 3) Counterfactual Inference: Counterfactual predictions from factual telemetry. The fitted model then predicts the follow three steps [2]. First, abduction estimates each UE’s counterfactual goodput used in the scheduling objective. residual state û(u) from factual telemetry using the assumed 1) Causal Discovery: In practice, the causal graph is not mechanism-specific noise rules. Its counterfactual interpretation given and must be recovered from data, a task known as causal assumes correctly specified parent sets and noise forms, discovery [5]. The DAG in Fig. 1 serves as the simulator accurate mechanism fits, and exogenous conditions that remain ground truth against which discovery is evaluated, and the fixed under the intervention. The inferred residuals approximate recovered graph supplies the parent sets used for structural these conditions but also contain model and estimation error. fitting and the causal order used for counterfactual prediction. Second, action replaces the PRB-allocation mechanism with We use the Peter–Clark (PC) algorithm [5] to learn an initial the assignment T = t. Finally, prediction evaluates the fitted causal skeleton from observable telemetry through conditional- GBM mechanisms in topological order, using updated parent independence testing. We restrict edge directions using protocol- values and applying each UE’s abduced residual according to layer and within-layer processing orders [14], [15] and orient the mechanism-specific rule. the remaining edges with the Direct Linear Non-Gaussian IV. S IMULATION R ESULTS AND A NALYSIS Acyclic Model (DirectLiNGAM) [17]. DirectLiNGAM exploits We evaluate the proposed framework in a single-cell non-Gaussianity to estimate a causal ordering under a linear downlink RAN simulation using Sionna 1.2 for the physical acyclic model. Since our mechanisms are nonlinear, we use this layer [20]1 . We use the 3GPP clustered delay line model ordering as a heuristic, followed by interventional refinement. C (CDL-C) channel at 2.6 GHz, with a fixed delay spread Regression on pooled observational and interventional data of 100 ns and a UE speed of 3 m/s [12]. The simulated proposes additional edges, which are admitted in significance MCS-specific BLER curves determine the highest-rate MCS order while preserving acyclicity. satisfying the BLER target at each UE’s effective SINR. Candidate edges are then refined using simulator intervenThe resulting MCS and BLER values feed the MAC and tions that hold selected mediators fixed. For an edge Vi → Vj , network-layer equations in Section II to compute service rate, the mediator set Mij contains variables between Vi and Vj queue backlog, packet loss, and goodput. Table I lists the key in protocol order that respond to interventions on Vi and are simulation parameters. either ancestors of Vj in the candidate graph or children of Vi Using data generated by combined PHY, MAC, and networkconnected by previously retained edges. These variables are layer simulation, we evaluate causal discovery on the 16 held fixed jointly, yielding the contrast observable telemetry variables, connected by 21 of the complete ∆ij = E[Vj | do(Vi = x1 ), do(Mij = m)] DAG’s 27 edges. The remaining six nodes are the latent (10) factors {ξsf , ξhw , ξtb } and noise terms {εγ , εT , ε3 }, which are − E[Vj | do(Vi = x0 ), do(Mij = m)] . 1 Code: https://github.com/THE-TRAIN-LAB/H-SCM-scheduling.

Fig. 2. Confounding analysis. (a) Goodput versus PRB allocation. (b) Full-PRB goodput estimates compared with interventional ground truth. (c) Goodput distributions. (d) Mean goodput by cell-load range.

unavailable in the assumed telemetry. Their outgoing edges are therefore excluded from discovery and its evaluation, while their effects remain present in the generated observations. Each discovery run uses 500 observational and 2940 interventional samples, with an additional 2000 paired UE realizations per mediator-testing contrast described in Section III-C1. Across 50 independent realizations, the proposed pipeline recovers 19 of the 21 observable edges without false positives, achieving precision 1.000, recall 0.905, and F1 score 0.950. The interventional regimes and the mediator-blocking contrasts set internal variables that only a simulator can control, which is a limitation of the discovery stage, whereas fitting for the counterfactual (CF) queries and held-out validation uses observational telemetry only. After recovering the causal graph, we fit the SCM according to (11) using 500 observational samples. Each GBM uses 200 trees, a maximum depth of 4, and a learning rate of 0.05. Prediction accuracy is measured by the coefficient of determination (R2 ), the fraction of ground-truth outcome variance explained by the predictions. For counterfactual prediction, residuals are additive for most mechanisms, multiplicative for the MAC service rate, and rescaled only in the traffic-dependent component for queue backlog. For CF-2 (hardware normalization), the MAC residual multiplier is set to one, representing nominal hardware under the assumed decomposition. On 5000 heldout UEs, the per-UE counterfactual predictions achieve R2 values of 0.979, 0.988, and 0.994 against the simulator’s perUE outcomes under do(T = 50), do(T = 100), and hardware normalization, respectively. Fig. 2 compares goodput predictions from naive ordinary least squares (OLS) regression of Y on PRB allocation T alone, confounding-adjusted OLS regression of Y on both T and cell load L, and the H-SCM, against ground-truth goodput from simulator interventions. In panels (a) and (b), naive OLS regression predicts a mean goodput of 49.20 Mbps at

T = 100, exceeding the simulator’s ground-truth interventional mean of 28.39 Mbps by 73.3%. This naive estimate omits L, unlike the conventional scheduler in (6). Including cell load L in the regression and averaging over the observed load distribution reduces the predicted mean goodput to 31.18 Mbps and the relative error to 9.8%. The recovered and reference graphs yield identical adjusted estimates because the two edges missed during causal discovery do not change the required adjustment variable L. In panel (b), averaging cross-fitted HSCM counterfactual predictions over the 500 observational UEs gives 27.64 Mbps, a 2.7% relative error against the simulator mean over 5000 held-out UEs. Panels (c) and (d) show the goodput distributions and responses across cell-load ranges. The mean interventional goodput increases from 2.42 to 14.12 and 28.39 Mbps as the allocation increases from 10 to 50 and 100 PRBs. However, fixing the allocation leaves the effect of load on link quality active, so performance still differs across load conditions. In Fig. 3, we show the results of three CF queries. CF-1 and (u) CF-3 evaluate the prediction Ŷt that constitutes the objective of (9) at the fixed allocation t = 100, whereas CF-2 evaluates hardware normalization as a model-based simulator diagnostic. CF-1, shown in panels (a) and (d), asks what goodput the UEs with moderate factual goodput, 5 ≤ Y ≤ 40 Mbps, would have attained had they received the full PRB allocation do(T = 100). This criterion selects 179 UEs, whose mean goodput increases from 19.53 to 29.04 Mbps, an improvement of 48.7%. Under full-PRB allocation, the mean increase in goodput relative to factual values is 5.65 Mbps for UEs in the most channel-limited residual range and 11.35 Mbps for those in the most favorable range. CF-2 asks what goodput the UEs underperforming their peers would have attained had their MAC residual multiplier been set to one. We select UEs whose MAC throughput is below 95% of the median among peers with the same MCS and PRB decile. Panel (b) shows the sorted per-UE goodput changes. Panel (e) shows the mean predicted goodput change across scaled MAC-residual ranges. Mean predicted goodput increases from 14.15 to 15.80 Mbps (11.7%), with gains concentrated among UEs with a negative scaled MAC residual, i.e., a residual multiplier below one; those with a positive residual lose goodput. CF-3, shown in panels (c) and (f), asks what goodput the UEs selected by the joint criteria of nominal SNR γ0 ∈ [10, 20] dB and factual goodput Y < 15 Mbps would have attained under do(T = 100). These criteria select 113 UEs, whose mean goodput increases from 3.55 to 5.52 Mbps, an improvement of 55.6%. However, UEs in the channel-limited residual range gain only 0.11 Mbps on average, compared with 6.88 Mbps for UEs in the favorable, PRB-limited range. These differences show that the mean gain alone cannot identify which UEs benefit from additional PRBs. Fig. 4 evaluates (9) for four UEs (K = 4) sharing the loaddependent PRB budget derived from (2), with at least 10 PRBs per UE. Training data comprise two scheduling epochs from 500 four-UE cells in the nominal environment, first under equal allocation and then a random feasible split. H-SCM uses 4000 unpaired rows. For scheduling, we fit the MAC mechanism to

V. C ONCLUSION In this paper, we have developed an H-SCM for crosslayer causal reasoning in wireless RANs. We have combined causal discovery, structural-equation fitting, and counterfactual inference to evaluate resource-allocation decisions. Simulation results have shown that accounting for cell-load confounding improves estimates of the effect of PRB allocation on goodput. Unit-level analysis has further shown that the same allocation can produce substantially different gains across UEs. Together, these results show that confounding adjustment and counterfactual inference address different needs in causality-aware RAN control. Future work will consider interventions that can be actuated on an operational RAN. We will also validate the recovered structure on Open RAN (O-RAN) traces. R EFERENCES Fig. 3. Counterfactual goodput gains: (a)–(c) sorted UE gains, where n is the number of UEs selected by each query, and (d)–(f) mean gains by inferred residual range. Dashed lines show the mean gain. Bar labels give UE counts.

Fig. 4. Scheduling performance across four environments: (a) loss relative to the ground-truth optimum and (b) realized cell goodput. Error bars show approximate 95% confidence intervals over 50 realizations.

µ/T using its remaining observable parents and reconstruct its prediction by multiplying by the candidate PRB allocation. An MLP with two hidden layers of 64 ReLU units each uses 2000 pairs to predict second-epoch goodput from full first-epoch telemetry and candidate allocation. Two reference policies are the same H-SCM without abduction, setting residuals to their population defaults, and the conventional scheduler of (6), fitted as a GBM of goodput on allocation and context x(u) using the same rows. All policies are tested without refitting across four environments over 50 realizations. Relative to nominal conditions, high load raises mean L from 0.50 to 0.71 and traffic-burstiness log-variance from 0.12 to 0.30. Poor channel lowers the maximum SNR from 28 to 16 dB, raises minimum Doppler/delay spread from 5 Hz/30 ns to 100 Hz/165 ns, and increases shadowing standard deviation from 7 to 9 dB. The diverse-UE environment increases hardware-factor standard deviation from 1.5 to 3, shadowing standard deviation from 7 to 10 dB, and traffic-burstiness log-variance from 0.12 to 0.40. Panel (a) reports mean relative loss in realized total cell goodput against the ground-truth optimum. Loss is 0.17– 0.35% for H-SCM, 0.8–1.8% without abduction, 2.0–2.9% for MLP, 8–10% for goodput-proportional allocation, and 19–42% for (6). Abduction reduces loss in every environment with identical mechanisms. Panel (b) confirms that H-SCM closely approaches ground-truth optimal goodput.

[1] M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia, “ColO-RAN: Developing machine learning-based xApps for open RAN closed-loop control on programmable experimental platforms,” IEEE Transactions on Mobile Computing, vol. 22, no. 10, pp. 5787–5800, 2023. [2] J. Pearl, Causality: Models, Reasoning, and Inference, 2nd ed. Cambridge, U.K.: Cambridge University Press, 2009. [3] E. Björnson and E. Jorswieck, “Optimal resource allocation in coordinated multi-cell systems,” Foundations and Trends in Communications and Information Theory, vol. 9, no. 2–3, pp. 113–381, 2013. [4] F. Lotfi, O. Semiari, and F. Afghah, “Evolutionary deep reinforcement learning for dynamic slice management in O-RAN,” in Proc. of the 2022 IEEE Globecom Workshops (GC Wkshps), Rio de Janeiro, Brazil, Dec. 2022, pp. 227–232. [5] J. Peters, D. Janzing, and B. Schölkopf, Elements of Causal Inference: Foundations and Learning Algorithms. Cambridge, MA, USA: MIT Press, 2017. [6] T. J. VanderWeele, “On the distinction between interaction and effect modification,” Epidemiology, vol. 20, no. 6, pp. 863–871, 2009. [7] C. K. Thomas, C. Chaccour, W. Saad, M. Debbah, and C. S. Hong, “Causal reasoning: Charting a revolutionary course for next-generation AI-native wireless networks,” IEEE Vehicular Technology Magazine, vol. 19, no. 1, pp. 16–31, 2024. [8] L. Xue, B. Zhang, Y. Ma, M. K. Marina, H. Yan, Y. Zhou, C. Y. Ip, S. Dhandapani, and J. Klosowski, “CausalTune: Causal learning based automated cellular RAN configuration tuning framework,” in Proc. of the ACM SIGCOMM 2026 Conference, Denver, CO, USA, Aug. 2026, pp. 242–262. [9] R. Prasad, M. Polese, and T. Melodia, “BLINC: Context-specific causal learning for automated RAN configuration,” arXiv preprint arXiv:2604.27084, 2026. [10] Q. Hou, S. Park, M. Zecchin, Y. Cai, G. Yu, and O. Simeone, “What if we had used a different app? Reliable counterfactual KPI analysis in wireless systems,” IEEE Transactions on Cognitive Communications and Networking, vol. 11, no. 5, pp. 3529–3543, 2025. [11] A. Farzaneh, S. D’Oro, and O. Simeone, “Should I have expressed a different intent? Counterfactual generation for LLM-based autonomous control,” arXiv preprint arXiv:2601.20090, 2026. [12] 3GPP, “Study on channel model for frequencies from 0.5 to 100 GHz,” 3rd Generation Partnership Project, Tech. Rep. TR 38.901 V17.0.0, Mar. 2022. [13] I. Siomina and D. Yuan, “Analysis of cell load coupling for LTE network planning and optimization,” IEEE Transactions on Wireless Communications, vol. 11, no. 6, pp. 2287–2297, 2012. [14] 3GPP, “NR; physical layer procedures for data,” 3rd Generation Partnership Project, Tech. Spec. TS 38.214 V19.4.0, Jun. 2026. [15] ——, “NR; medium access control (MAC) protocol specification,” 3rd Generation Partnership Project, Tech. Spec. TS 38.321 V19.3.0, Jun. 2026. [16] W. Whitt, “The queueing network analyzer,” Bell System Technical Journal, vol. 62, no. 9, pp. 2779–2815, Nov. 1983. [17] S. Shimizu, T. Inazumi, Y. Sogawa, A. Hyvärinen, Y. Kawahara, T. Washio, P. O. Hoyer, and K. Bollen, “DirectLiNGAM: A direct method for learning a linear non-Gaussian structural equation model,” Journal of Machine Learning Research, vol. 12, pp. 1225–1248, 2011. [18] J. H. Friedman, “Greedy function approximation: A gradient boosting machine,” The Annals of Statistics, vol. 29, no. 5, pp. 1189–1232, 2001. [19] V. Chernozhukov, D. Chetverikov, M. Demirer, E. Duflo, C. Hansen, W. Newey, and J. Robins, “Double/debiased machine learning for treatment and structural parameters,” The Econometrics Journal, vol. 21, no. 1, pp. C1–C68, 2018. [20] J. Hoydis, S. Cammerer, F. Ait Aoudia, A. Vem, N. Binder, G. Marcus, and A. Keller, “Sionna: An open-source library for next-generation physical layer research,” arXiv preprint arXiv:2203.11854, 2022.

Related documents

• From Confounding to Counterfactuals: A Cross-Layer Causal Model for RAN Scheduling#1536463
• A Quantitative Measure Of Fairness And Discrimination For Resource Allocation In Shared Computer Systems#6601
• Resource Allocation, Control, and Accounting for the Use of Network Resources#136690
• Scheduling in Multi-Hop Wireless Networks With Deadlines#120460
• Budgeted Task-Aware Acquisition of Dynamic Networks#973483