{"id":"193d7d5b-0a4d-4208-9024-7550c624b4f7","arxiv_id":"2504.14085","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"A multi-armed bandit that discretizes and searches non-uniform preamble access probabilities comes within about 5% of an SQP-optimized allocation for high-priority RACH throughput, but its reward scaling depends on the unknown device count.","lead":"This paper uses a multi-armed bandit to choose non-uniform random access probabilities for high- and low-priority machine devices in a cellular RACH, aiming to boost high-priority throughput while holding low-priority throughput above a threshold. The headline claim is that the bandit approach needs no knowledge of the device count, but the reward scaling step appears to require exactly that count.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reward scaling in Section IV-C requires nh, so the MAB as specified cannot be deployed without UE counts; the abstract's headline claim is therefore unsupported.","rationale":"The reader's weakest assumption correctly locates a real gap: Section IV-C's reward scaling factor depends on nh, and the conclusion's future-work statement admits the BS does not initially know nh, contradicting the abstract. I partially agree with the reader. However, I do not treat this as an inherent impossibility: because the scaling factor is a single positive constant applied to every reward, replacing it with any other positive constant preserves the ranking of actions and hence the MAB's returned action. That observation means the flaw is in the manuscript's specification, not in the underlying idea. Still, the paper as written does not say the scale is arbitrary, does not provide a calibration procedure that avoids nh, and its Algorithm 1 cannot be instantiated by a BS that knows only observed patterns. Since the central claim 'does not require the BS to know the number of UEs' is not supported by the stated algorithm, the reader's REJECT verdict stands. I would not soften to CONDITIONAL without the authors explicitly removing or reworking the scaling step and reconciling the abstract with the conclusion.","tokens_in":20726,"tokens_out":27142,"duration_ms":254628,"concrete_test":"Re-run the MAB of Section V-D2 (Tables VII/VIII) under two reward scalings: (i) the Section IV-C divisor computed from the true nh, and (ii) a fixed positive divisor independent of nh, e.g., scale(µh,T) = µh,T. Compare the returned actions and the resulting µh,T and µl,T for every M. If all rows return identical actions, the nh-dependence is removable and the paper needs only a clarification that scaling is an arbitrary positive constant; if any row changes, the algorithm as specified depends on nh and the headline claim is false.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract claims the MAB does not require the BS to know nh/nl. But Algorithm 1 computes the reward via scale(µh,T), and Section IV-C defines the scaling factor as the expected µh when ph, pl are set to (9)–(10). For those vectors, L-UEs are all sent to RB M, so H-UEs contend only over the other M−1 RBs and the divisor equals nh (1 − 1/(M−1))^{nh−1}, an explicit function of nh. A BS that does not know nh cannot evaluate this divisor, and Algorithm 1 has no nh input. The paper's own conclusion lists 'where the BS does not initially know the number of H-UEs and L-UEs' as future work, directly contradicting the abstract. This is not a remote edge case: the scaling step sits inside every Q-value update, so the algorithm as written and simulated requires the very quantity the headline says is unnecessary. It is true that if the divisor is any fixed positive constant, the argmax over actions is preserved because the scaling is action-independent; but the paper neither states nor exploits that fact, and the reported implementation is not reproducible by a BS that knows only observed access patterns.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies a two-priority random access channel (RACH) in which high-priority and low-priority UEs choose preambles according to non-uniform probability vectors. It defines a pattern-based throughput model, formulates a constrained optimization problem for the access probabilities, solves it with SQP, and then proposes a multi-armed bandit (MAB) approach with cross-entropy updates to approximate the solution. Two action-space constructions are considered: a discretized action space with circular-shift redundancy removal, and a compact action space (CAS) built from a precomputed lookup table. The paper's headline claim is that the MAB solution is near-optimal and scalable and does not require the BS to know the number of UEs.","tokens_in":20930,"tokens_out":14098,"duration_ms":134872,"significance":"The idea of optimizing non-uniform preamble-selection probabilities for prioritized RACH access is interesting, and the optimization results in Tables IV and V are plausible under the stated multinomial model. The MAB formulation with cross-entropy updates and the action-space reduction are useful engineering ideas. However, the central no-knowledge claim is not supported by the manuscript as written: the reward-scaling step requires knowledge of nh, the CAS evaluation is circular, and the simulation appendix is not reproducible as printed. If the no-knowledge claim could be substantiated, the contribution would be significant for mMTC deployment; in its current form, the paper needs substantial revision.","major_comments":[{"comment":"The reward-scaling step divides µh,T by the value of µh obtained when ph and pl are set to (9)-(10). Under the multinomial model of Section II, this divisor equals nh(1 - 1/(M-1))^{nh-1}, which is an explicit function of the unknown nh. Algorithm 1 has no input for nh, and the experiments in Tables VII and VIII use known values nh=4, nl=5. A BS that does not know nh cannot evaluate the scaling factor, so the algorithm as specified is not deployable in the setting claimed in the abstract. The paper's own conclusion lists 'where the BS does not initially know the number of H-UEs and L-UEs' as future work, directly contradicting the abstract. Because scaling is applied inside every Q-value update, this is a load-bearing issue. Please either modify the scaling to be computable without nh, add an estimation stage, or remove the no-knowledge claim from the abstract and contributions.","section":"Section IV-C, Algorithm 1, Eqs. (9)-(10)"},{"comment":"The compact action space is built by solving the same optimization problem from Section III-B for each (nh,nl) pair, and the numerical evaluation in Section V-D3 tests the MAB on the same load pairs that were used to populate the lookup table, including (nh,nl)=(4,5). The true optimum is therefore contained in the action set by construction, and the MAB is essentially selecting among precomputed optima rather than discovering near-optimal probabilities. This makes the reported CAS performance a tautology rather than evidence for generalization. Please test on held-out load pairs, or reposition the CAS as a table-lookup method with no generalization claim.","section":"Section IV-B2, Section V-D3"},{"comment":"The probability P(π) is the foundation of the throughput metrics µh and µl in Eqs. (1)-(2) and of the optimization problem (4), but its derivation is deferred to the authors' own companion paper [34], and the recursive backtracking algorithm for enumerating Π is omitted. As a result, the optimization results cannot be verified from the manuscript alone. Please either include the derivation of the factors in Eq. (3) and the backtracking procedure, or clearly state that the model is inherited from [34] and provide enough detail for an independent check.","section":"Section III-A, Eq. (3)"},{"comment":"Algorithm 5 is not reproducible as printed. If C is the (nh+nl) by T matrix described in lines 3-5, then 'for c in C' iterates over rows (individual UEs), not over RACH slots, so the per-slot event labeling described in lines 12-21 is applied to per-UE time histories rather than to slot-wise access patterns. Furthermore, lines 14-21 identify whether a successful UE is H or L by comparing the slot index j with the RB index i, which is not meaningful; the H/L distinction should come from the row's position in C. Please provide the corrected pseudocode or the actual simulation code used to generate Tables II-VIII.","section":"Appendix B, Algorithm 5"},{"comment":"The uniform-access benchmark values in Table II appear inconsistent with the model of Section II. For nh=4, nl=5, and M=3, Eqs. (1)-(3) give µh = 4(2/3)^8 ≈ 0.156 and µl = 5(2/3)^8 ≈ 0.195, whereas Table II reports 0.31 and 0.62. For M=5 the model gives µh ≈ 0.671 and µl ≈ 0.839, versus the reported 0.40 and 0.53. Since Table II is used as the uniform benchmark in later comparisons, please reconcile these values with the model or the simulation, and clarify whether the reported numbers are means, medians, or values from a single realization.","section":"Table II"}],"minor_comments":[{"comment":"The text says 'the reward is µl,T multiplied by the penalty factor' when the L-UE constraint is violated, but Eq. (6) defines the penalized reward as ρ µh,T. Please align the text with the equation.","section":"Section IV-A, after Eq. (6)"},{"comment":"There are several typographical errors: 'Futhermore' in the Introduction, 'surpases' in the Conclusion, and 'utilitez' in Appendix A. The sentence about the simulation hardware is also duplicated in Section V.","section":"Throughout"},{"comment":"In panels (d) of Figures 3 and 4, the legend shows 'µh,Tµh,T' instead of 'µh,Tµl,T'. Please correct the legends.","section":"Figures 3 and 4"},{"comment":"The text reports a maximum achievable H-UE throughput of 1.2282 for nh=4, nl=5, but Table V gives µh = 1.28 for the same load at M=5. Please clarify whether 1.2282 is a scaled value or belongs to a different configuration.","section":"Section V-D4"}],"recommendation":"major_revision","confidential_remarks":"The main concern is that the abstract's headline claim ('does not require the BS to know the number of UEs') is contradicted by the paper's own conclusion and by the reward-scaling implementation, which needs nh. In addition, the key probability derivation is placed in a self-cited companion paper [34], and the CAS evaluation is circular. I would ask the authors to provide the simulation code and to revise either the algorithm or the claims substantially before the paper can be considered further."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi [name],\n\nQuick read of arXiv:2504.14085. The core idea is legitimately new: non-uniform preamble selection for a two-priority RACH, with an SQP optimization and a cross-entropy MAB to find the access probabilities. The throughput formulas are simple, and the reported numbers are consistent with them. The improvement over uniform access and ACB is real in the simulated regimes (roughly doubling H-UE throughput, with the L-UE constraint met). That part deserves credit.\n\nThe soft spot is the headline claim. The abstract says the MAB “does not require the BS to know the number of UEs in the network.” But the reward scaling in Section IV-C divides µh,T by the expected µh when ph, pl are set to (9)–(10). For those vectors, L-UEs are all sent to one RB, so H-UEs contend over M−1 RBs and the divisor is nh(1−1/(M−1))^(nh−1) — an explicit function of nh. Algorithm 1 has no nh input. A BS that truly doesn’t know nh cannot compute its own rewards. The paper’s own conclusion lists “where the BS does not initially know the number of H-UEs and L-UEs” as future work, which directly contradicts the abstract. That’s not a minor wording issue; it’s the main selling point of the RL approach.\n\nOther issues are more standard: the P(π) derivation is deferred to a self-cited companion paper, which makes the throughput model hard to verify standalone. The CAS lookup table is built from the same optimization and tested on the same (nh,nl) pairs, so that variant is optimistic by construction. No error bars, though T=1000 slots makes the variance small.\n\nOn balance, I’d send this to peer review rather than desk-reject. The core non-uniform access idea is interesting and the experiments are plausible, but the “no knowledge of UE counts” claim needs to be walked back or fixed. A reasonable revision would either drop the claim, use an estimated nh in the scaling, or note that the scaling factor is action-independent and therefore argmax-preserving even if unknown. The paper is worth a serious referee’s time.\n\nBest,","headline":"A genuinely new non-uniform access scheme for two-priority RACH, with plausible results, but the headline claim that the MAB needs no knowledge of UE counts fails because the reward scaling in Section IV-C requires nh.","tokens_in":21497,"tokens_out":2594,"would_cite":false,"duration_ms":23315,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Multi-armed bandits can tune random-access preamble choices to favor high-priority users, reaching within 5% of the constrained optimum.","keywords":["massive machine-type communications (mMTC)","multi-armed bandits (MAB)","random access channel (RACH)","reinforcement learning (RL)","access probability optimization","non-uniform preamble selection","priority access","network load estimation"],"falsifier":"Run the multi-armed bandit with the reward scaler replaced by a normalization computed only from observed throughputs (or set to 1), give the base station no access to the numbers of high- and low-priority devices, and compare the returned high-priority throughput with the sequential-quadratic-programming optimum for a fixed load such as $n_h=4$, $n_l=5$, $M=6$, $\\gamma=0.4$; a gap beyond 5% or a violation of the low-priority constraint would refute the no-count claim as written.","tokens_in":20461,"feed_emoji":"📶","tokens_out":8821,"duration_ms":75407,"temperature":0.7,"pith_summary":"The paper's central claim is that non-uniform preamble selection in the cellular random access channel lets a base station favor high-priority devices while keeping low-priority service above a threshold, and that this choice can be learned online. It builds a two-priority model in which the base station sees each slot as a pattern of single high-priority successes, single low-priority successes, empty resource blocks, and collisions, and from these observable patterns it can score candidate access-probability vectors. The exact optimum is formulated as a constrained nonlinear problem, and a multi-armed bandit with a discretized action space is proposed to approximate it; the reported results put the bandit within 5% of the constrained optimum and show it adapting to a mid-run change in the number of devices. If the claims hold, the practical upshot is a scalable way to prioritize machine-type traffic without redesigning the RACH protocol.","feed_headline":"Bandit-tuned RACH preambles hit within 5% of optimal priority access","feed_subtitle":"A base station can favor high-priority devices by tuning preamble probabilities from observed successes and collisions.","key_machinery":"The load-bearing object is the access pattern: each RACH slot is summarized as a word over four event types—single high-priority success, single low-priority success, empty resource block, and collision—and expected throughputs are written as sums over the probabilities of these patterns. That model turns the problem into a constrained nonlinear optimization over the two probability vectors, and then into a multi-armed bandit in which each arm is a discretized pair of probability vectors, redundant circular shifts are removed offline, rewards are the observed high-priority throughput (penalized when low-priority throughput falls below the threshold), and cross-entropy updates concentrate the sampling distribution on high-reward arms without visiting every action.","core_discovery":"The paper's central discovery, stated in its own terms, is that the optimal preamble-selection probabilities in a two-priority RACH are non-uniform: high-priority UEs spread across most resource blocks while low-priority UEs concentrate on a few, and this allocation beats both uniform selection and access class barring on high-priority throughput under the same low-priority floor. The multi-armed bandit formulation reaches within 5% of the constrained optimum (Tables VII and VIII) and, unlike the sequential-quadratic-programming optimum, runs in roughly constant reported wall-clock time as the number of resource blocks grows from 3 to 6 in the constrained case; the compact action space variant cuts that time further and also produces a network-load estimate whose mean absolute error decreases with runs. The paper also reports that the bandit adapts when the network load changes mid-simulation.","pith_inferences":["The paper's conclusion lists 'the BS does not initially know the number of H-UEs and L-UEs' as future work, while the abstract claims no such knowledge is needed; this asymmetry suggests the count-free claim should be read as holding only after a load-estimation step has supplied the reward-scaling constant.","A straightforward testable extension is to replace the count-dependent reward scaler with a normalizer computed only from observed throughputs (for example, the batch maximum of $\\mu_{h,T}$, or $M$ itself) and re-run Tables VII and VIII; if the within-5% gap survives, the no-count claim becomes empirically supported.","Because the model only needs the four event types, the same bandit framing should extend to more than two priority classes or to traffic with non-saturated arrivals, though the paper does not report those cases."],"forward_implications":["A base station can tune preamble selection probabilities for each priority class from observed access patterns alone, without changing the RACH protocol.","Non-uniform allocation gives higher high-priority throughput than uniform allocation or access class barring while holding low-priority throughput at the required threshold.","The multi-armed bandit evaluates only a small curated set of probability pairs, so the approach remains feasible when the number of resource blocks grows large enough to make exact optimization impractical.","The compact action space version can return an estimate of the number of high- and low-priority devices as a byproduct of the lookup table.","When the device population changes mid-operation, the learned probabilities re-converge to the new operating point."],"supporting_citations":[{"why":"Supplies the derivation of the access-pattern probability factors used in the throughput formulas (1)–(3).","marker":"[34]"},{"why":"Provides the cross-entropy multi-armed bandit update that lets the learner concentrate on elite actions without sweeping the full action space.","marker":"[44]"},{"why":"Gives the sequential quadratic programming solver that computes the constrained optimum used as the accuracy benchmark.","marker":"[35]"},{"why":"Defines RACH throughput as the average number of successful UEs per slot, the objective being optimized.","marker":"[13]"},{"why":"Establishes the uniform-allocation benchmark for single-priority random access that motivates the base-case allocation.","marker":"[26]"},{"why":"Supports the modeling assumption that the base station cannot count colliding UEs, so only the four event types are observable.","marker":"[23]"},{"why":"Supports choosing cross-entropy over other bandit algorithms for large action spaces.","marker":"[45]"}],"fun_headline_variants":["Bandit-tuned RACH preambles near-optimal for two-priority access","Multi-armed bandits optimize RACH access probabilities","Reinforcement learning tunes RACH preambles for priority UEs","Two-priority RACH: bandits hit 95% of optimal access rate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reward-scaling step divides the measured high-priority throughput by a value computed from the numbers of high- and low-priority users, so the claim that the base station needs no knowledge of those counts holds only if a separate count or estimate is available.","fun_headline_variants_meta":{"raw":{"variants":["Bandit-tuned RACH preambles near-optimal for two-priority access","Multi-armed bandits optimize RACH access probabilities","Reinforcement learning tunes RACH preambles for priority UEs","Two-priority RACH: bandits hit 95% of optimal access rate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000378,"raw_usage":{"total_tokens":1990,"prompt_tokens":903,"completion_tokens":1087,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":1007}},"tokens_in":519,"tokens_out":1087,"duration_ms":9309,"temperature":1.0,"reasoning_tokens":1007,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:57:20.570547+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the multi-armed bandit with the reward scaler replaced by a normalization computed only from observed throughputs (or set to 1), give the base station no access to the numbers of high- and low-priority devices, and compare the returned high-priority throughput with the sequential-quadratic-programming optimum for a fixed load such as $n_h=4$, $n_l=5$, $M=6$, $\\gamma=0.4$; a gap beyond 5% or a violation of the low-priority constraint would refute the no-count claim as written.","supporting_citations":[{"cited_title":"Load estimation in a two-priority mMTC random access channel,","cited_arxiv_id":null,"evidence_quote":"Supplies the derivation of the access-pattern probability factors used in the throughput formulas (1)–(3)."},{"cited_title":"Cemab: A cross-entropy- based method for large-scale multi-armed bandits,","cited_arxiv_id":null,"evidence_quote":"Provides the cross-entropy multi-armed bandit update that lets the learner concentrate on elite actions without sweeping the full action space."},{"cited_title":"NLPQL: A FORTRAN subroutine solving constrained nonlinear programming problems,","cited_arxiv_id":null,"evidence_quote":"Gives the sequential quadratic programming solver that computes the constrained optimum used as the accuracy benchmark."},{"cited_title":"Priority- based load-adaptive preamble separation random access for QoS- differentiated services in 5G networks,","cited_arxiv_id":null,"evidence_quote":"Defines RACH throughput as the average number of successful UEs per slot, the objective being optimized."},{"cited_title":"PRADA: Prioritized random access with dynamic access barring for MTC in 3GPP LTE-A networks,","cited_arxiv_id":null,"evidence_quote":"Establishes the uniform-allocation benchmark for single-priority random access that motivates the base-case allocation."},{"cited_title":"PPO-based PDACB traffic control scheme for massive IoV communications,","cited_arxiv_id":null,"evidence_quote":"Supports the modeling assumption that the base station cannot count colliding UEs, so only the four event types are observable."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports choosing cross-entropy over other bandit algorithms for large action spaces."}],"review_version":1}