{"id":"ee214334-9e4e-43b9-8941-166c77e9762e","arxiv_id":"2501.01281","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A DRL-based block coordinate descent method for fluid antenna positioning improves simulated communication rates in multi-target ISAC over fixed-position antennas.","lead":"This letter proposes a deep reinforcement learning framework to jointly tune beamforming and antenna positions in a fluid antenna integrated sensing and communication system with multiple radar targets. Simulations show higher communication rates than fixed-position antennas, but the results omit sensing-gain verification and error bars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (12) has sign-reversed penalty terms: the sensing and power penalties vanish exactly when constraints are violated, so the DRL objective does not match the constrained problem in (9), and the reported rate gains may reflect constraint violation.","rationale":"The paper's contribution is an algorithmic recipe plus simulation evidence. There is no code release, no formal verification, and no ablation of the reward components, so the only check on correctness is the consistency between the described constrained problem, the DRL reward, and the reported numbers. Eq. (12) is the bridge between the DDPG policy and constraints (9e)/(9f), and as written it pushes the agent toward exactly the states that violate those constraints. This is more specific than the reader's concern: it is not merely that alpha1 and Gamma are unreported, but that the stated reward structure is inconsistent with the stated objective. The missing sensing-gain reporting compounds the problem, because even a correct reward needs empirical verification that varphi >= Gamma at the reported rates. I do not see reason to question the authors' intent; a sign error or a results-reporting gap is the most plausible explanation. Therefore the appropriate outcome remains the reader's CONDITIONAL verdict, with the condition broadened to include correcting Eq. (12) and reporting the actual sensing gains and reward weights. I would not move to unconditional acceptance, and I would not reject out of hand because the framework is plausible and the flaw is fixable. The concrete test determines which of the two readings is true.","tokens_in":6706,"tokens_out":9221,"duration_ms":90945,"concrete_test":"Implement Eq. (12) literally and evaluate two mock states: (A) varphi = 0 for all targets, Tr(U) = Pmax; (B) varphi = 2Gamma, Tr(U) = Pmax/2. If reward(A) > reward(B), the printed objective rewards constraint violation. Then re-run the DDPG-BCD training with corrected penalty terms (alpha1 * max(0, Gamma - varphi^(m)) and alpha2 * max(0, Tr(U) - Pmax)) using the same environment and hyperparameters, and compare the N=8, K=3 curve in Fig. 2. Also report achieved varphi(p^(m)) against Gamma and the selected alpha1, alpha2, alpha3, Gamma. If the published curve is reproduced only under the printed (inverted) reward, the rates are obtained under the wrong objective; if reproduced under corrected penalties, the sign is a typo and the required revision is to state the correction and report the sensing gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B1c defines the reward that is the sole training signal for antenna positions. As printed, Eq. (12) subtracts alpha1 * max(0, varphi(p^(m)) - Gamma) and alpha2 * max(0, Pmax - Tr(U)). These terms are zero when varphi < Gamma or Tr(U) > Pmax, i.e., precisely when (9f) or (9e) is violated, and positive when the constraints are satisfied with margin. That is the reverse of the paper's own description ('penalizes violations ... ensuring that the sensing gain ... remains greater than Gamma'). If training used this reward, the DDPG agent is not optimizing the constrained problem (9); the higher rates versus FPA in Figs. 2-3 could simply reflect dropped sensing and power requirements. If, instead, Eq. (12) is a typo and the implemented reward used max(0, Gamma - varphi) and max(0, Tr(U) - Pmax), the manuscript still omits alpha1, alpha2, alpha3, Gamma, and all achieved sensing gains, so there is no evidence that the reported operating points satisfy (9f). In either reading, the central claim that FAS-DRL 'balances' sensing and communication is not supported by the presented material.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers an integrated sensing and communication (ISAC) system in which a dual-functional base station equipped with fluid antennas (FAs) communicates with a user terminal and senses multiple radar targets. It formulates a joint transmit-covariance and antenna-position optimization problem (Problem (9)) and proposes a block coordinate descent (BCD) framework: for fixed antenna positions, the transmit covariance matrix U is obtained by solving a convex optimization problem (10) with CVX and, if needed, Gaussian randomization; for fixed U, the antenna positions are optimized by a deep deterministic policy gradient (DDPG) agent whose state includes positions and beamforming features, whose action is incremental antenna displacements, and whose reward is given by Eq. (12). Numerical results in Figs. 2 and 3 claim higher average and maximum communication rates than a fixed-position-antenna (FPA) baseline, and the authors further claim scalability to more users and targets. The central claim is that continuous FAS repositioning via DRL provides a practical way to balance communication and sensing in multi-target ISAC.","tokens_in":6999,"tokens_out":5247,"duration_ms":57714,"significance":"The topic is timely: FAS and ISAC are both active 6G research areas, and the specific problem of jointly optimizing beamforming and continuous antenna positions for multi-target sensing is not yet thoroughly explored. The convex beamforming subproblem is formulated and solved in a standard, correct manner, and the use of DDPG to handle the non-convex position subproblem is a plausible design choice that, if validated, could be useful for practical deployment. However, the evidence presented in the manuscript is currently insufficient to support the central claim. The reward function printed in Eq. (12) has penalty terms whose signs are reversed relative to the constraints, no sensing-gain or power-consumption values are reported to verify feasibility of the operating points, and the claimed scalability to multiple users is not backed by a multi-user formulation. These issues are fixable, but the manuscript as written does not substantiate the claimed communication-sensing balance.","major_comments":[{"comment":"The reward function as printed is sign-reversed with respect to the constraints of Problem (9). The first penalty term is −α1 Σ_m max(0, ϖ(p^(m)) − Γ), which is zero when the sensing gain violates the constraint ϖ(p^(m)) ≥ Γ and positive when the constraint is satisfied with margin; the second penalty term is −α2 max(0, Pmax − Tr(U)), which is zero when the power constraint Tr(U) ≤ Pmax is violated and positive when it is satisfied. Thus the penalties vanish exactly on constraint violation, so an agent maximizing this reward is not optimizing the constrained problem (9) and may be encouraged to violate the sensing and power constraints. This contradicts the paper's own description in Section III-B1c, which states that the second term 'penalizes violations of the required sensing gain constraint.' If the implemented reward used the opposite signs, the manuscript must state this explicitly and correct Eq. (12); if the implementation matches the printed formula, the rate gains over FPA in Figs. 2 and 3 cannot be interpreted as evidence of a valid constrained solution.","section":"Section III-B1c, Eq. (12)"},{"comment":"The manuscript never reports the values of the reward weights α1, α2, α3, the sensing gain threshold Γ, or any achieved sensing gain ϖ(p^(k)) or transmit power Tr(U) at the reported operating points. These quantities are essential because the claimed 'balance' between communication and sensing rests on the assertion that constraint (9f) and (9e) are actually satisfied. Without verifying that the DRL policy meets the sensing-gain requirements, the higher average rates in Figs. 2 and 3 may simply reflect the agent's freedom to violate the sensing and power constraints. Please add a table or figure showing the achieved sensing gains and transmit powers (or their average/feasibility fractions) for the reported configurations, together with the exact values of α1, α2, α3, and Γ used in training.","section":"Section IV, Tables/Figs. 2 and 3"},{"comment":"The system model, channel model, and optimization problem (9) are all formulated for a single user terminal (M = 1). Yet Eq. (12) contains a sum over m = 1, ..., M without defining M, and Fig. 3 reports results for M = 3 user terminals. The multi-user extension is mentioned only in a footnote and in the discussion of Fig. 3; no multi-user channel vector, rate expression, beamforming constraints, or problem formulation is provided. Consequently, the paper's scalability claim for 'additional users' is not mathematically supported. Either provide the full multi-user formulation and explain how the BCD-DRL algorithm is adapted to it, or restrict the scalability claims to the number of targets K rather than the number of users.","section":"Section II and Section III-B, multi-user claim"}],"minor_comments":[{"comment":"The paper gives some network and training parameters (buffer size 10000, mini-batch 64, target update τ = 0.001) but omits other essential DDPG hyperparameters, including the learning rates for actor and critic, the discount factor, the Ornstein-Uhlenbeck coefficients ξ and ς, and the number of training episodes. Please report the hyperparameter setting to make the results reproducible.","section":"Section IV, DDPG hyperparameters"},{"comment":"The quantities 'Average Communication Rate' and 'Maximum Communication Rate' are not precisely defined; it is unclear whether the average is over random channel realizations, training seeds, or both, and what the maximum is taken over. Please define these metrics and include error bars or confidence intervals across multiple random seeds, as DRL training is sensitive to seed variability.","section":"Figs. 2 and 3"},{"comment":"There is an unbalanced parenthesis in the first max term of Eq. (12): 'max(0, ϖ(p^(m)) − Γ' is missing a closing parenthesis. Also, the variable M is used in Eq. (12) but is not defined in the problem formulation; if it denotes the number of user terminals, a definition should be given before first use.","section":"Section III-B1c, Eq. (12)"},{"comment":"There are several typographical errors that should be corrected, including 'receiever' in Section II-B, 'constrants' in Section III-B1c, and 'trade-sff' in the title of reference [9]. Additionally, the paper only compares against an FPA baseline; a comparison with at least one other movable-antenna or FAS position-optimization baseline (e.g., the iterative method of [8] or a random search) would strengthen the claim of superiority.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The decisive issue is the sign reversal in Eq. (12). If the implementation actually follows the printed reward, the numerical results do not support the constrained optimization claim and the paper would need rejection; if it is a typo, the manuscript can be repaired by correcting Eq. (12), reporting the missing hyperparameters and constraint-satisfaction metrics, and either formulating the multi-user case or narrowing the scalability claims. Since the error is local and fixable within the scope of a revision, I recommend major revision rather than rejection. Please also verify that the described BCD loop is actually what was run: the paper does not specify how DDPG training interacts with the CVX step, and no pseudocode or convergence criterion is provided, which affects reproducibility and the 'real-time decision-making' claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The actual novelty here is modest but real: continuous fluid-antenna positioning via DDPG inside a BCD loop for multi-target ISAC, where prior work used discrete port selection or single-target settings. The problem formulation is clean, the convex beamforming subproblem is correctly relaxed and solved, and the simulation curves show a consistent rate advantage over fixed-position antennas as the number of antennas and targets grows. That part I believe.\n\nThe soft spots are not fatal to the idea, but they are bigger than the paper admits. First, Eq. (12) as printed subtracts alpha1 * max(0, varphi(p^(m)) - Gamma) and alpha2 * max(0, Pmax - Tr(U)). Those terms are zero exactly when the sensing or power constraint is violated, and positive when the constraint is satisfied with margin. That is the reverse of a penalty. The surrounding text describes the opposite intent, so this may be a typo, but as written the DRL objective does not match problem (9). If the implementation used the printed reward, the rate gains over the FPA baseline could simply reflect broken constraints. If it used the intended reward, the manuscript still never reports achieved sensing gains, the weights alpha1, alpha2, alpha3, or Gamma, so there is no evidence that the operating points satisfy (9f). Either way, the central claim of \"balancing\" sensing and communication is not demonstrated.\n\nThere are also no error bars or seed variability, and the only baseline is FPA, which makes the scalability claims look better than the evidence supports. These are all fixable in revision. The paper deserves a serious referee because the approach is plausible and the subfield is active, but it should not be accepted until the reward function is corrected and the sensing constraints are actually verified. I would not cite it in its current form.","headline":"A plausible but under-verified DRL recipe for FAS-ISAC; Eq. (12) as printed has sign-reversed penalties, so the central performance claims are not supported as written.","tokens_in":7534,"tokens_out":4434,"would_cite":false,"duration_ms":40463,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A DRL-BCD framework that alternates convex beamforming with DDPG-based antenna positioning achieves higher communication rates than fixed-position antennas in multi-target ISAC, and scales to more users and targets.","keywords":["fluid antenna systems","integrated sensing and communication","deep reinforcement learning","DDPG","block coordinate descent","beamforming","antenna positioning","multi-target sensing"],"falsifier":"Inspect the converged DDPG policy on a fixed set of channel realizations at a given SNR: record the achieved sensing gain $\\varpi(p^{(k)})$ for each target $k$. If a non-negligible fraction of targets fall below the threshold $\\Gamma$ while the reported communication rate remains high, then the reported rate is partly an artifact of the reward penalty, not a genuine sensing-communication trade-off. The same check applies to the rank-one reconstructed beamformer from Gaussian randomization, which may violate the sensing constraint (9f) even when the relaxed convex solution satisfies it.","tokens_in":6527,"feed_emoji":"📡","tokens_out":7817,"duration_ms":65718,"temperature":0.7,"pith_summary":"This paper tries to establish that a dual-function base station with fluid antennas can serve one communication user and sense multiple targets better if antenna positions are optimized by a deep-reinforcement-learning policy rather than fixed. The authors propose a block coordinate descent framework that alternates a convex optimization step for the transmit covariance matrix with a deep deterministic policy gradient (DDPG) step that moves the fluid antennas continuously inside a small region. In simulation, the resulting communication rates exceed those of a fixed-position antenna baseline even when the baseline serves only one target, and the approach extends to multiple users and targets. The intended payoff is a practical, real-time method for balancing sensing and communication in ISAC systems without resorting to intractable joint optimization.","feed_headline":"DRL-placed fluid antennas beat fixed arrays for ISAC","feed_subtitle":"A BCD-DRL scheme balances communication and multi-target sensing, scaling to more users.","key_machinery":"The load-bearing object is the BCD decomposition of the original problem: for fixed antenna positions, the beamforming subproblem is convex in the transmit covariance matrix $U$ and is solved with a convex solver after relaxing the rank-one constraint, with Gaussian randomization restoring a rank-one solution; for fixed $U$, the position subproblem is recast as a Markov decision process whose state is the antenna coordinate vector plus three beamforming features (trace, largest eigenvalue, and mean eigenvalue of $U$), whose action is incremental position updates bounded within $\\pm A/2$ for $A = 4\\lambda$, and whose reward is the communication rate minus penalties for sensing-gain violations, power violations, and movement distance. The DDPG actor-critic, trained with Ornstein-Uhlenbeck exploration noise, learns a deterministic continuous positioning policy. This machinery converts an intractable joint optimization into alternating convex solves and policy-gradient updates that can execute in real time once trained.","core_discovery":"The central claim is that the joint non-convex problem of beamforming and antenna placement in a FAS-aided ISAC system can be decomposed into two alternating blocks: a convex transmit-covariance optimization, solved under a relaxed rank-one constraint, and an antenna-position Markov decision process solved by DDPG. Under a far-field channel model, the paper reports that the proposed DRL-BCD algorithm achieves higher average communication rates than a fixed-position antenna baseline across SNR values, with the gap persisting when multiple sensing targets are added. The paper also reports that adding more users increases maximum rates (for example, 27.6% higher at 30 dB with 12 antennas compared with a single user), and that the DRL policy generalizes to these larger setups.","pith_inferences":["A natural next step is to verify sensing constraint satisfaction explicitly; if the penalty weights are large enough, the same reward-shaping recipe could serve as a generic solver for other continuous-placement multi-objective wireless problems.","The far-field constant-AoA assumption holds only for small movement regions; testing with near-field or environment-dependent path parameters would reveal how much of the gain survives realistic propagation.","Replacing the hand-tuned penalty weights with a constrained reinforcement learning update (for example a Lagrangian method) would remove the need to choose $\\alpha_1,\\alpha_2,\\alpha_3$ and would make the claimed sensing-communication balance more robust."],"forward_implications":["In multi-target ISAC, communication rates improve versus fixed-position antennas even when the sensing workload grows, so antenna mobility is a degree of freedom worth exploiting.","The BCD-DRL framework scales to more users and targets without redesigning the algorithm, because the DDPG state and action dimensions grow linearly.","After training, antenna positioning decisions can be made in real time, unlike iterative alternating optimization, making the approach usable in fast-changing channel conditions.","Continuous antenna repositioning outperforms discrete port selection, suggesting that fine-grained placement freedom is valuable."],"supporting_citations":[{"why":"Supplies the far-field movable-antenna channel model with constant AoAs/AoDs that the paper adopts.","marker":"[6]"},{"why":"Provides the single-target FAS-ISAC baseline that this work extends to multiple sensing targets.","marker":"[8]"},{"why":"Represents the discrete-port selection approach the paper contrasts with continuous antenna positioning.","marker":"[9]"},{"why":"Contributes the deep-reinforcement-learning treatment of FAS-ISAC that this paper adapts to continuous positions.","marker":"[10]"},{"why":"The convex solver used in the beamforming subproblem of the BCD loop.","marker":"[11]"},{"why":"Defines fluid antenna systems and the spatial flexibility that motivates the positioning optimization.","marker":"[5]"}],"fun_headline_variants":["DRL antenna positioning boosts ISAC rates and scales to many users","FAS meets DRL: joint beamforming and antenna placement for ISAC","Deep RL improves fluid antenna placement for integrated sensing and communication","DRL-BCD antenna positioning beats fixed arrays in multi-target ISAC","Scalable DRL antenna placement for multi-user FAS-ISAC systems"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's central claim depends on the untested assumption that the penalty term in the reward function really forces the DDPG policy to meet the required sensing strength $\\varpi(p^{(k)}) \\ge \\Gamma$ for every target; the paper never reports the achieved sensing strengths or the penalty weights $\\alpha_1, \\alpha_2, \\alpha_3$ and threshold $\\Gamma$ it used.","fun_headline_variants_meta":{"raw":{"variants":["DRL antenna positioning boosts ISAC rates and scales to many users","FAS meets DRL: joint beamforming and antenna placement for ISAC","Deep RL improves fluid antenna placement for integrated sensing and communication","DRL-BCD antenna positioning beats fixed arrays in multi-target ISAC","Scalable DRL antenna placement for multi-user FAS-ISAC systems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000734,"raw_usage":{"total_tokens":3218,"prompt_tokens":819,"completion_tokens":2399,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":435,"completion_tokens_details":{"reasoning_tokens":2306}},"tokens_in":435,"tokens_out":2399,"duration_ms":16072,"temperature":1.0,"reasoning_tokens":2306,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:30:28.791035+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the converged DDPG policy on a fixed set of channel realizations at a given SNR: record the achieved sensing gain $\\varpi(p^{(k)})$ for each target $k$. If a non-negligible fraction of targets fall below the threshold $\\Gamma$ while the reported communication rate remains high, then the reported rate is partly an artifact of the reward penalty, not a genuine sensing-communication trade-off. The same check applies to the rank-one reconstructed beamformer from Gaussian randomization, which may violate the sensing constraint (9f) even when the relaxed convex solution satisfies it.","supporting_citations":[{"cited_title":"MIMO capacity characterization for movable antenna systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the far-field movable-antenna channel model with constant AoAs/AoDs that the paper adopts."},{"cited_title":"Shifting the ISAC Trade-Off with Fluid Antenna Systems","cited_arxiv_id":"2405.05715","evidence_quote":"Represents the discrete-port selection approach the paper contrasts with continuous antenna positioning."},{"cited_title":"CVX: Matlab software for disciplined convex programming, version 2.1,","cited_arxiv_id":null,"evidence_quote":"The convex solver used in the beamforming subproblem of the BCD loop."},{"cited_title":"Fluid antenna systems,","cited_arxiv_id":null,"evidence_quote":"Defines fluid antenna systems and the spatial flexibility that motivates the positioning optimization."}],"review_version":1}