{"id":"f3ba0866-8616-47da-96b1-3a0d9dac16d8","arxiv_id":"2501.15203","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A Soft Actor-Critic agent that adaptively tunes Particle Swarm Optimization parameters is reported to reduce simulated IIoT edge offloading cost by 28% versus plain PSO.","lead":"This paper describes a simulation where a reinforcement learning agent tunes the parameters of a particle swarm optimizer to lower task offloading costs in industrial internet of things edge computing. The authors report a 28 percent cost reduction versus plain particle swarm optimization, but without code link, error bars, or independent baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 28.38% improvement may be an artifact of the cost function: a task assigned to a server with insufficient RAM is scored with zero computational cost, making infeasible offloads cheap unless an unstated hard constraint prevents them.","rationale":"Reading the paper in good faith, the central claim is that SAC-driven adaptive control of PSO hyperparameters yields materially lower offloading costs. The only evidence offered is the simulated cost comparison in Section IV, so the most load-bearing condition is that the cost function correctly evaluates feasible offloading solutions. The equation for C_j creates a zero-cost loophole: a device assigned to a server with insufficient RAM pays no computational cost, and because m is 1000 times larger than n, this zeroing can dominate the objective. Neither the PSO nor the APSO update procedure is described as enforcing RAM feasibility, so the optimizer is free to prefer invalid assignments. This is not an external-validity concern; it threatens the internal validity of the reported numbers. If a meaningful fraction of the final assignments in the APSO-SAC runs are RAM-infeasible, then the 58.5 average best cost is not a valid cost for the constrained task-offloading problem. The reader's weakest_assumption was about generalizability to real IIoT environments, but I see a more immediate problem inside the simulation itself, so I disagree with the reader's identification of the weakest point. The overall verdict remains REJECT because the central claim is not supported as written; no verdict adjustment is needed, though the reason is sharper than the reader's rationale.","tokens_in":4914,"tokens_out":3950,"duration_ms":38726,"concrete_test":"Re-run the Section IV experiments with RAM feasibility enforced as a hard constraint: during pbest/gbest updates, either discard any candidate assignment in which a device's RAM requirement exceeds its assigned server's RAM, or add a large finite penalty for such violations. Also instrument the current implementation to report the fraction of devices assigned to insufficient-RAM servers for PSO, APSO, and APSO-SAC. If the infeasible-assignment fraction is nonzero, or if APSO-SAC's cost advantage shrinks materially under the hard constraint, the headline improvement is an artifact of the zero-cost loophole rather than a genuine algorithmic gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A defines the total cost as the sum over devices of [m * C_j + n * T_j], with C_j = ServerCost_sj * T_j if RAMRequirement_j <= ServerRAM_sj and C_j = 0 otherwise. The accompanying text claims this conditional 'ensures a device is only assigned to a server if the server has sufficient RAM, but the formula does no such thing: it merely charges zero computational cost for an invalid assignment while still counting T_j. Because m = 10 and n = 0.01, zeroing C_j removes the dominant cost term. The PSO/APSO update equations (Eq. 1) are unconstrained in the written method; no mechanism is described for filtering RAM-infeasible particles during pbest/gbest updates or final selection. Consequently, the optimizer can lower total cost by exploiting infeasible assignments, and the 28.38% advantage of APSO-SAC over PSO (Section IV-C, average best costs 58.5 versus 81.69) may reflect a greater tendency to discover such invalid assignments rather than better task offloading. This is an internal validity issue: even within the paper's own simulated setup, the best-cost comparison is not trustworthy unless RAM feasibility is enforced as a hard constraint and not as a zero-cost penalty.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes APSO-SAC, a hybrid that uses a Soft Actor-Critic (SAC) reinforcement learning agent to adapt the acceleration coefficients of Particle Swarm Optimization (PSO) for task offloading in an IIoT/MEC simulation. The authors define a total-cost objective in Section IV-A that weights computational cost and latency, then compare the average best cost over 10 runs for standard PSO, Adaptive PSO, and APSO-SAC. They report average best costs of 81.69, 76.39, and 58.5, respectively, claiming a 28.38% improvement for APSO-SAC over PSO, and conclude that SAC-controlled parameter adaptation produces better offloading decisions.","tokens_in":5173,"tokens_out":4561,"duration_ms":45929,"significance":"The general idea of using RL to adapt PSO hyperparameters is plausible and potentially useful for dynamic MEC task offloading, and a demonstrated improvement over PSO/APSO would be of practical interest. The paper also promises public code and data on GitHub, which is a positive step toward reproducibility. However, the contribution is currently not established: the central empirical claim rests on a single synthetic scenario, ten runs with no variance or statistical tests, hand-set cost weights, and a cost function that may reward RAM-infeasible assignments. As written, the reported gains are not trustworthy evidence of better task offloading, and the level of algorithmic detail is too low for the method to be independently verified.","major_comments":[{"comment":"The central empirical claim is supported only by the average best cost of ten runs, with no standard deviation, confidence interval, or significance test. The phrase 'consistently achieved the lowest best costs' is not justified by a single point estimate per method. The authors should report the full distribution of results over multiple independent seeds, provide statistical comparisons, and ideally vary the number of devices, servers, and task parameters to show the result is not an artifact of one randomly generated scenario.","section":"Section IV-A/IV-C"},{"comment":"The experimental comparison includes only standard PSO, APSO, and APSO-SAC on one synthetic scenario with fixed ranges and fixed weights. The conclusion that the method 'outperforms traditional PSO and other RL-integrated algorithms' is not supported, since no other RL-integrated baseline is evaluated. The external validity of the result also requires sensitivity analysis over the cost weights m and n and over the parameter ranges in Table II, none of which is provided.","section":"Section IV-B/IV-C"}],"minor_comments":[{"comment":"The text says 'Figure 4 shows the result of running an instance of Adaptive PSO' in the Adaptive PSO subsection, but Figure 4 is captioned 'APSOSAC'; the figure reference appears to be wrong and should be corrected.","section":"Section IV-B.2 and IV-B.3"},{"comment":"The GitHub availability statement gives no repository URL or identifier, so the claimed public code and datasets cannot actually be located by a reader.","section":"General"},{"comment":"There are numerous surface errors, including 'in static environment's', 'it's insensitivity', inconsistent use of commas in equations, and several dangling formulation issues in the total-cost description; these should be cleaned up.","section":"General"},{"comment":"The abstract and conclusion repeatedly claim 'optimal' task offloading, but no optimality guarantee or comparison to an optimal/exhaustive solution is provided; the claim should be softened to 'improved' unless proven otherwise.","section":"Conclusion"}],"recommendation":"reject","confidential_remarks":"The main result is contaminated by the RAM-feasibility loophole in the cost function and by the absence of statistical evidence; these are not local presentation issues. I would need to see a substantially revised evaluation with hard feasibility constraints, proper baselines, and significance testing before the paper could be considered further, and I would not invite a revision at this stage given the fundamental nature of the flaws."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central claim is that SAC-tuned PSO cuts offloading cost by 28% over standard PSO, averaging 58.5 vs 81.69. On reading, I think that advantage is more likely an artifact of the cost function than a real optimization win. The stress-test note holds up: in Section IV-A, C_j is ServerCost_sj * T_j when RAM fits and 0 otherwise, and the text says this “ensures” a device is only assigned to a server with sufficient RAM. It does no such thing. Zeroing the dominant cost term (m=10) makes infeasible assignments artificially cheap, and the PSO update equations are unconstrained—no pbest/gbest filtering or final-solution check is described. So the optimizer can lower total cost by assigning tasks to servers with too little RAM, and the 28% gap may just mean the SAC-controlled variant found more of those invalid assignments. That is an internal validity problem, not just a transferability concern.\n\nWhat the paper does well: combining SAC as an adaptive controller for PSO hyperparameters is a reasonable idea, and the authors are honest about APSO's 110% runtime increase. The simulation setup is clearly described and the parameter ranges are sensible. The problem is a real one.\n\nSoft spots beyond the cost-function flaw: no variance or significance tests on the 10 runs, no seed reported, no working code repository despite promising one, and a figure/heading error (the APSOSAC section references Figure 4 as “Adaptive PSO”). The conclusion claims superiority over “other RL-integrated algorithms” that were never compared. The writing is rough in places but understandable.\n\nThis is a paper where the load-bearing result collapses on inspection. It is not a desk reject for being frivolous—the idea is plausible and the experimental design is almost right—but it needs major revision before it can support the reported numbers. The authors need to enforce RAM feasibility as a hard constraint (or honestly penalize infeasibility), rerun the experiments, and report distributions.\n\nWho is this for? Someone using it as a cautionary example of how a constraint encoded as a zero-cost penalty can sink an optimization comparison. I would not bring it to reading group, and I would not cite it. I would not send it to peer review in its current form: the central evidence is compromised, and the paper is not important enough to justify referee time on a known-broken comparison. If the authors fix the constraint and rerun, the revised version might deserve a look.","headline":"The reported 28% cost improvement is probably an artifact of a cost function that zeroes the dominant term for RAM-infeasible assignments; the paper's central claim is not trustworthy as written.","tokens_in":5652,"tokens_out":2442,"would_cite":false,"duration_ms":23224,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A reinforcement-learning agent that tunes particle swarm optimization reduces IIoT task offloading cost by 28.4% in simulated edge-computing tests.","keywords":["task offloading","mobile edge computing","Industrial Internet of Things","particle swarm optimization","adaptive PSO","soft actor-critic","reinforcement learning","resource allocation"],"falsifier":"Run the same optimizers on a real or recorded set of factory workloads with fixed device and server specifications; if APSO-SAC's average best cost over repeated runs is not below both PSO and APSO, or if its margin disappears when the cost and latency weights are varied, the central claim is falsified.","tokens_in":4723,"feed_emoji":"🏭","tokens_out":7840,"duration_ms":61962,"temperature":0.7,"pith_summary":"This paper proposes a hybrid optimizer that lets a reinforcement-learning agent govern the adaptive parameters of particle swarm optimization for making task offloading decisions in mobile edge computing. The authors aim to establish that this hybrid, called APSO-SAC, produces the lowest total offloading cost among the methods tested: an average best cost of 58.5 versus 81.69 for standard PSO and 76.39 for adaptive PSO alone, a 28.38% reduction relative to PSO. The wider point is that a learned controller can replace manual, problem-specific tuning of swarm hyperparameters while keeping runtime low, which matters for industrial IoT settings where task offloading decisions must adapt to changing conditions.","feed_headline":"RL-tuned PSO cuts IIoT offloading cost by 28.4%","feed_subtitle":"A swarm optimizer with a reinforcement-learning controller beats plain PSO in simulated offloading.","key_machinery":"The mechanism is the velocity update of PSO, $v_i(t+1) = w_t v_i(t) + c_1 U(a,b)(pbest_i - x_i) + c_2 U(a,b)(gbest - x_i)$, whose coefficients are normally fixed or manually adapted. The paper replaces manual adaptation with a SAC controller that, at each observation of the swarm's performance, outputs adjustments to the acceleration coefficients, aiming to maximize the expected return plus an entropy bonus that encourages exploration. The total-offloading cost function $\\sum_j [m C_j + n T_j]$, with $m=10$ and $n=0.01$, measures both computational cost $C_j$ and latency $T_j$ and is the quantity the optimizer minimizes.","core_discovery":"On the paper's own terms, the central discovery is that inserting a Soft Actor-Critic policy between the swarm state and the PSO velocity update converts a static optimizer into a self-tuning one: the SAC agent observes the swarm's behavior and adjusts the acceleration coefficients $c_1$ and $c_2$, and the result is a consistently lower best cost over ten simulation runs. The measured values are 58.5 for APSO-SAC versus 81.69 for PSO and 76.39 for APSO, with runtime comparable to PSO rather than the roughly 110% runtime increase that APSO alone suffers. The authors interpret this as evidence that RL-controlled adaptation improves both exploration and exploitation in the offloading search, yielding more efficient assignments of 250 devices to 20 MEC servers in the simulated environment.","pith_inferences":["The cost weights $m=10$ and $n=0.01$ put far more weight on computational cost than on latency, so the reported improvement may be driven mostly by cheaper server choices; a reader should check the latency component separately before assuming end-to-end delay also improved.","Devices and servers are randomly generated within fixed ranges and placements are static, so the paper's framing about adapting to human interactions and environmental changes is not yet demonstrated; a dynamic workload trace would be the natural follow-up test.","The SAC controller adjusts only the acceleration coefficients $c_1$ and $c_2$; extending it to the inertia weight $w_t$, population size, or the number of iterations could plausibly yield further gains."],"forward_implications":["If the result is correct, an RL-tuned swarm optimizer can reduce average best offloading cost by roughly 28% over standard PSO in simulated MEC environments.","The approach keeps runtime near the level of PSO, avoiding APSO's 110% runtime penalty while still improving on APSO's cost.","Because SAC is relatively insensitive to hyperparameter tuning, the method can be deployed without the domain-specific manual adaptation that plain APSO requires.","The same optimizer architecture could apply to other weighted-sum assignment problems in edge and fog computing, not only the specific device-server scenario tested."],"supporting_citations":[{"why":"defines mobile edge computing and motivates placing computation closer to devices, the setting the optimizer targets.","marker":"[1]"},{"why":"provides the PSO-based task offloading approach whose average best cost serves as the main baseline.","marker":"[5]"},{"why":"supplies an improved PSO variant for IoT offloading that the adaptive mechanism extends.","marker":"[6]"},{"why":"represents the deep RL approach to MEC resource management that the paper contrasts with its hybrid.","marker":"[7]"},{"why":"introduces adaptive PSO, whose evolutionary-factor state transitions the SAC controller is designed to mimic.","marker":"[9]"}],"fun_headline_variants":["RL-tuned PSO beats plain PSO in IIoT offloading","SAC-controlled swarm optimizer cuts edge costs by 28%","Hybrid SAC-PSO improves IIoT task offloading efficiency","Self-tuning PSO via RL lowers offloading cost in IIoT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The random parameter ranges of Table II and the chosen cost weights $m=10$, $n=0.01$ stand in for real IIoT conditions, so the reported cost improvement may not carry over to actual factory deployments if those ranges or weights are unrepresentative.","fun_headline_variants_meta":{"raw":{"variants":["RL-tuned PSO beats plain PSO in IIoT offloading","SAC-controlled swarm optimizer cuts edge costs by 28%","Hybrid SAC-PSO improves IIoT task offloading efficiency","Self-tuning PSO via RL lowers offloading cost in IIoT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1397,"prompt_tokens":815,"completion_tokens":582,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":431,"completion_tokens_details":{"reasoning_tokens":508}},"tokens_in":431,"tokens_out":582,"duration_ms":5655,"temperature":1.0,"reasoning_tokens":508,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:29:58.188774+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same optimizers on a real or recorded set of factory workloads with fixed device and server specifications; if APSO-SAC's average best cost over repeated runs is not below both PSO and APSO, or if its margin disappears when the cost and latency weights are varied, the central claim is falsified.","supporting_citations":[{"cited_title":"Mobile edge com- puting: A survey,","cited_arxiv_id":null,"evidence_quote":"defines mobile edge computing and motivates placing computation closer to devices, the setting the optimizer targets."},{"cited_title":"Efficient task offloading using particle swarm optimization algorithm in edge computing for industrial internet of things,","cited_arxiv_id":null,"evidence_quote":"provides the PSO-based task offloading approach whose average best cost serves as the main baseline."},{"cited_title":"Computation offloading strategy for iot using improved particle swarm algorithm in edge computing,","cited_arxiv_id":null,"evidence_quote":"supplies an improved PSO variant for IoT offloading that the adaptive mechanism extends."},{"cited_title":"Deep reinforcement learning-based dynamic resource management for mobile edge computing in industrial internet of things,","cited_arxiv_id":null,"evidence_quote":"represents the deep RL approach to MEC resource management that the paper contrasts with its hybrid."},{"cited_title":"Adaptive particle swarm optimization,","cited_arxiv_id":null,"evidence_quote":"introduces adaptive PSO, whose evolutionary-factor state transitions the SAC controller is designed to mimic."}],"review_version":1}