{"id":"f626252d-db90-461e-a024-2464c9acf79f","arxiv_id":"2505.04272","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A deep reinforcement learning offloading policy combined with a knapsack-based channel allocator lowers simulated delay-energy cost for mobile edge computing with dependent tasks and user mobility.","lead":"The paper couples a deep Q-network offloading policy with a dynamic programming channel allocator for mobile edge computing networks where users move and tasks depend on one another. It reports lower delay-energy cost than several comparison methods in simulation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DCA's knapsack cost ignores which subchannels are assigned, so the joint optimizer does not actually minimize the Eq. (6)-(10) cost; the reported gains may depend on a random tie-break rather than on the claimed channel-aware allocation.","rationale":"The reader's weakest assumption and my load-bearing concern are the same: DCA allocates counts of subchannels through a grouped knapsack but assigns the specific subchannels randomly, even though the paper's rate expression in Eq. (6) makes the cost depend on the identities of the subchannels through heterogeneous gains and co-channel interference. This is not a minor implementation detail: it means the algorithm does not solve the channel-allocation subproblem as formulated, so the 'joint optimization' claim is not internally supported. The concern is empirical as well as theoretical, because the D3QN reward is computed after the random assignment, making the episode costs in Figs. 5-9 random variables whose reported values could depend on the particular random draw and on data sizes rather than on the quality of the allocation policy. The proposed concrete check, replacing the random assignment with a deterministic channel-aware assignment, would settle whether the claimed lowest cost survives. I agree with the reader's conditional verdict: the paper's engineering heuristic may still perform well, but the current evidence does not establish that the channel-allocation component is actually optimizing the stated cost. I would not move the verdict to reject because the empirical comparison could still be valid in expectation, and the proposed test is straightforward to run.","tokens_in":19798,"tokens_out":5967,"duration_ms":65460,"concrete_test":"Rerun the Fig. 6 bandwidth sweep with a deterministic subchannel assignment within DCA: after the DP fixes the count z_n per terminal, assign specific subchannels to maximize the actual sum-rate (or minimize the true communication cost in Eq. (9)) for each ES, e.g., by greedy or Hungarian matching using the gains h_{n,m,k}[t] and interference I_{n,k}[t] from Eq. (6), instead of a uniform random draw. Also rerun the original random assignment 10 times with different seeds. If the TOICA cost changes materially (e.g., >5%) or the ordering against SEG+DCA/ON+DCA changes, the claimed advantage is an artifact of the random tie-break; if the cost is stable and still lowest, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that TOICA achieves the least delay-energy cost because D3QN offloading is jointly optimized with DCA channel allocation. The load-bearing step is DCA's claim to solve the channel-allocation subproblem. In Section III-A, DCA builds a grouped knapsack whose item value v_{n,z} is the 'resulting communication cost' of giving terminal n exactly z subchannels, then, after the DP returns the counts z, 'different subchannels are randomly assigned to MTs according to the quantity requirements.' This is internally inconsistent with the channel model in Eq. (6), where the rate depends on the specific channel gain h_{n,m,k}[t] and on the interference pattern I_{n,k}[t] created by which MTs share each subchannel. Two allocations with the same count z but different subchannels can have very different communication times and therefore very different values of d_{n,i,m} in Eq. (9). The DP therefore minimizes a proxy (a count-based cost) rather than the actual objective (14); the 'optimal' count vector is optimal only for an unspecified averaging assumption that the paper does not state or verify. Since the actual cost used in the D3QN reward is computed after a random draw of specific subchannels, the reported episode costs are random variables. The comparison in Fig. 6 may therefore reflect a particular random realization and the specific data sizes, not a property of the proposed channel-allocation method. This is the soft spot: the paper's strongest empirical claim rests on an allocation step that does not solve the channel-allocation problem it defines.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies joint task offloading and subchannel allocation in a multi-user multi-server MEC network with terminal mobility and dependent tasks modeled as DAGs. It formulates a long-term delay-energy weighted cost minimization problem (14), decouples task dependencies via a priority rule based on average cost, and proposes TOICA, which combines a dynamic programming-based channel allocation algorithm (DCA) built on a grouped knapsack model and a D3QN-based offloading decision algorithm (DTO). The authors claim that TOICA achieves the lowest delay-energy cost among the compared baselines in simulations. The manuscript includes a system model with co-channel interference and I/O interference, a complexity analysis of DCA, and simulation results over varying bandwidth, edge server capacity, and network scale.","tokens_in":20169,"tokens_out":3766,"duration_ms":39429,"significance":"If the claims hold, the paper would make a useful contribution to MEC resource management by jointly addressing spatial dynamics, task dependencies, and channel-aware offloading with a learning-based method. The explicit modeling of co-channel interference, I/O interference, and DAG scheduling, together with the integrated D3QN/knapsack architecture, is a plausible and relevant direction. The paper also provides a complexity analysis and evaluates adaptability to different weights and network sizes. However, the central algorithmic claim rests on the correctness of DCA, and the current presentation leaves several load-bearing gaps in that algorithm's definition and in the empirical validation.","major_comments":[{"comment":"The DCA algorithm's random assignment of specific subchannels is internally inconsistent with the channel model in Eq. (6). The rate r_up_{n,m}[t] depends on the specific channel gain h_{n,m,k}[t] and on the interference pattern created by which terminals share each subchannel, yet the knapsack DP in Section III-A optimizes only over the number z of subchannels per terminal, and then 'different subchannels are randomly assigned to MTs according to the quantity requirements.' Two allocations with the same count vector can have very different communication times and therefore very different costs under Eqs. (7)-(9). As written, the DP minimizes a count-based proxy, not the actual objective (14), so the reported performance gain may be an artifact of the random tie-breaking rather than of the claimed channel-aware allocation. The authors should either define v_{n,z} using an explicit averaging assumption over subchannel realizations and validate that assumption, or modify DCA to optimize over specific subchannel assignments.","section":"Section III-A, Eq. (6), Algorithm 1"},{"comment":"The item value v_{n,z} in the grouped knapsack formulation is never defined. The text calls it 'the resulting communication cost' but provides no formula connecting z to the data load, channel gains, or interference terms of Section II. Without an explicit expression for v_{n,z}, the DP's optimality claim for the channel allocation subproblem is not verifiable, and the mapping from the original problem to the GKP instance is incomplete. A precise definition of v_{n,z} is essential for the paper's central claim that DCA optimally allocates channels for a given offloading decision.","section":"Section III-A, Eq. (15)"},{"comment":"The NP-hardness argument is logically reversed. The paper states that 'our channel allocation subproblem can be directly mapped to a GKP instance in polynomial time' and concludes that problem (14) inherits NP-hardness. Reducing the subproblem to GKP shows that the subproblem is no harder than GKP, not that GKP (or the original problem) is NP-hard. To establish NP-hardness, the authors need a polynomial-time reduction from a known NP-hard problem (e.g., GKP) to problem (14). This should be corrected either by supplying a proper reduction or by softening the claim to 'the problem is combinatorial and presumably NP-hard, motivating a heuristic approach.'","section":"Section III-A, NP-hardness discussion"},{"comment":"The simulation results are reported without error bars, confidence intervals, or multiple random seeds. Figures 5 and 9 show single convergence traces, and Figures 6-8 compare costs without any statistical measure. Since the DCA step involves random subchannel assignment and the environment includes Rayleigh fading and random mobility, the reported costs are random variables, and the claim that 'TOICA conducts the least cost among all the algorithms' may not be robust to different realizations. The authors should run multiple independent trials and report mean plus/minus standard deviation (or confidence intervals), and preferably use a paired statistical test for the cost comparisons.","section":"Section IV, Figs. 5-9"}],"minor_comments":[{"comment":"The term 'duel DQN' is a typographical error; it should be 'dueling DQN' (and similarly in the phrase 'duel DQN' in Section III-B).","section":"Abstract and Section III-B"},{"comment":"The definition of D[t] in (14a) uses a max over tasks, while the objective sums D[t] over t. Since each task is assumed to finish within one slot, it would be clearer to define D[t] as the maximum completion time among tasks executed in slot t and to state explicitly that only one task per terminal is active per slot.","section":"Eq. (14a) and Eq. (14)"},{"comment":"The summations over t in Eqs. (4) and (5) are confusing because on_{i,m}[t] is an indicator of offloading at slot t, but the model states each task completes in a single slot. It would help to clarify whether the sum is over the slot in which the task is processed, and to simplify the notation accordingly.","section":"Eqs. (4)-(5)"},{"comment":"In Algorithm 1, the backtracking step sets 'j←K' and uses a loop over 'reverse(Z_n)' but the condition and the break logic are not fully specified; also the final line 'Assign subchannels according to alloc' hides the random assignment step that is critical to the cost evaluation. The pseudocode should be tightened to match the described procedure.","section":"Algorithm 1 and Section IV"},{"comment":"The y-axis of Fig. 5 is 'cumulative rewards,' but the text describes cost comparisons; consider labeling the axes consistently and explaining the relationship between the accumulated reward and the cost metric used in the other figures.","section":"Fig. 5 and Fig. 9"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the overall direction is reasonable, but the internal inconsistency in DCA and the undefined v_{n,z} are load-bearing issues that require careful revision. The simulation methodology also needs to meet the standard of statistical rigor for this venue. I do not see evidence of intentional misrepresentation, but the current presentation overstates the optimality of DCA and the strength of the empirical claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a reasonable engineering contribution: it combines dependent-task DAG scheduling, a grouped-knapsack channel allocator, and D3QN for offloading, plus an I/O-interference model for edge servers. The specific combination is new in the MEC literature, and the simulation campaign is broader than many in this subfield (bandwidth, compute capability, delay-energy weight, network scale). If the algorithm performs as claimed, it is a useful addition for MEC practitioners.\n\nThe main soft spot is the DCA. The knapsack item value v_{n,z} is called the \"resulting communication cost\" of giving terminal n exactly z subchannels, but after the DP returns counts, the actual subchannels are randomly assigned to terminals. That is inconsistent with the channel model in (6), where the rate depends on the particular gains h_{n,m,k}[t] and on the interference pattern I_{n,k}[t] created by which terminals share each subchannel. So the DP is minimizing a count-based proxy, not the actual objective, and the reported episode costs are random variables after the random draw. This is a legitimate concern, and the performance comparison in Fig. 6 may partly reflect favorable random realizations. It is not fatal—one can read DCA as a heuristic—but the paper should either define v_{n,z} explicitly as an expected cost over subchannels or replace the random assignment with a simple matching. At minimum, the authors should report statistics over multiple random assignments.\n\nOther issues are minor. The NP-hardness argument is a sketch, not a formal reduction. The simulations have no error bars or multiple seeds, which weakens the strength of \"the proposed TOICA conducts the least cost\" claim. The \"first to comprehensively consider\" statement is a bit strong given that all components exist in the cited literature, but the citation pattern itself is fine—the papers cited are relevant and the combination is genuinely new.\n\nOverall the paper deserves peer review. The model is plausible, the algorithm is clearly described, and the empirical story is coherent. The DCA inconsistency is a fixable flaw that a good referee will catch; it does not by itself invalidate the work. I would send it out and ask for a major revision focused on the channel-allocation step and on reporting variance.","headline":"Solid incremental MEC paper with a real but fixable channel-allocation inconsistency; deserves peer review.","tokens_in":20695,"tokens_out":2518,"would_cite":false,"duration_ms":27551,"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":"TOICA couples D3QN offloading with knapsack channel allocation to cut delay-energy cost under mobility and task dependencies.","keywords":["mobile edge computing","task offloading","channel allocation","deep reinforcement learning","D3QN","grouped knapsack problem","task dependencies","delay-energy trade-off"],"falsifier":"Replace DCA's final random assignment of specific subchannels with a channel-state-aware assignment under the same knapsack counts and rerun the simulations; if the cost drops materially, the current allocator is not optimal under the paper's own heterogeneous channel model in equation (6).","tokens_in":19609,"feed_emoji":"⚡","tokens_out":6832,"duration_ms":62206,"temperature":0.7,"pith_summary":"This paper tackles task offloading and channel allocation together in a mobile edge computing network where terminals move and applications have dependent tasks. It claims that jointly optimizing the two decisions—offloading choices via a D3QN agent and subchannel counts via a grouped-knapsack dynamic program—yields lower long-term delay-energy cost than existing dependency-aware offloading schemes, greedy offloading, and nearest-server offloading. The paper's contribution is a decomposition that keeps the two decisions coupled through the reward signal rather than solving them as one hard combinatorial problem. If right, the work matters because it offers a real-time-capable, scalable way to allocate communication and computation resources under both spatial mobility and temporal task dependencies.","feed_headline":"MEC offloading algorithm cuts delay-energy cost in dynamic networks","feed_subtitle":"Joint D3QN offloading and knapsack channel allocation beat four baselines in moving MEC simulations.","key_machinery":"The engine is a two-level decomposition. First, tasks are ordered by a priority rule: each task's priority is its estimated average processing cost plus, for non-exit tasks, the largest priority among its successors; ordering by this priority turns the directed acyclic graph into a topological sequence. Second, given offloading decisions, each edge server solves a grouped knapsack problem—groups are terminals, items are possible subchannel counts, weights are counts, values are communication costs—via dynamic programming, returning the number of subchannels per terminal. Third, each terminal runs a D3QN agent whose action is local execution or one of the edge servers, state is position plus task data and computation size, and reward is $\\tanh(\\text{cost}_{n,i,0}-\\text{cost}_{n,i,m'})$, where the offloaded cost is computed from the allocation returned by the channel allocator. The channel allocation result enters the reward, which is how channel allocation and offloading are coupled.","core_discovery":"The paper claims that in a multi-user, multi-server MEC network with moving terminals and dependent application tasks, the joint problem of where to offload each task and how to assign uplink subchannels can be split into two coupled pieces. A priority rule based on estimated average processing cost turns each terminal's task graph into a topological execution order. Given the offloading choices, each edge server allocates subchannel counts by solving a grouped knapsack problem with dynamic programming; the resulting allocation feeds into the reward of a D3QN agent that chooses, for each task, between local execution and one of the edge servers. The paper reports that this TOICA algorithm converges to the lowest delay-energy cost among the compared schemes, tracks the trade-off coefficient $\\omega$ between delay and energy, and remains convergent in networks with 3, 6, and 9 edge servers.","pith_inferences":["Editorial inference: since DCA randomizes which subchannels a terminal receives after fixing the count, a frequency-aware assignment within the same knapsack solution would likely reduce cost further; this is testable by replacing the random draw with a greedy assignment and rerunning the comparison.","Editorial inference: the same decomposition—priority ordering for dependency constraints, a knapsack for resource-count decisions, and a DRL agent for discrete choices—could carry over to other resource dimensions, such as computation shares or cache slices, where count decisions separate from item identity.","Editorial inference: the 36.8 ms-per-task figure comes from a laptop CPU inside the simulator; measuring the online decision phase on representative edge hardware would show whether the 10–100 ms real-time target is met outside the simulated environment."],"forward_implications":["If TOICA works as claimed, MEC operators can tune one weight $\\omega$ to shift system behavior between delay-critical and energy-critical applications without changing the algorithm.","Larger bandwidth and finer subchannel granularity (16 vs 8 subchannels) lower the weighted cost for the proposed scheme.","Increasing edge-server computing capacity reduces the cost of all schemes, with TOICA remaining the lowest in the reported experiments.","The policy can be trained offline and then deployed online, with simulated per-task decisions around 36.8 ms, inside the 10–100 ms range cited for autonomous driving and real-time video analytics.","In larger networks with 3, 6, and 9 edge servers, the training still converges, and adding servers under fixed terminal load reduces the weighted cost."],"supporting_citations":[{"why":"supplies the grouped-knapsack reformulation for subchannel allocation","marker":"[10]"},{"why":"provides the D3QN algorithm used for offloading decisions","marker":"[19]"},{"why":"supplies the priority-based approach for decoupling dependent tasks","marker":"[20]"},{"why":"provides the HRROGA comparison baseline and simulation parameter settings","marker":"[6]"},{"why":"provides the SEG comparison baseline and simulation parameter settings","marker":"[8]"},{"why":"motivates the VM I/O interference model in edge-server computation time","marker":"[26]"},{"why":"establishes grouped knapsack as NP-hard, supporting the NP-hardness argument","marker":"[27]"},{"why":"supplies the DQN mechanism with target network and experience replay that D3QN builds on","marker":"[28]"}],"fun_headline_variants":["D3QN and knapsack team up to beat dynamic MEC baselines","Joint offload and channel scheme cuts delay-energy in moving MEC","Spatial-temporal MEC offloading solved by D3QN and knapsack","Knapsack channel allocation + D3QN offloading beats dynamic baselines","Hybrid D3QN-knapsack scheme lowers MEC delay-energy cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that, once the number of subchannels for each terminal is fixed, the identity of those subchannels does not affect the result—the algorithm assigns them randomly—even though the model gives different subchannels different channel gains and interference levels.","fun_headline_variants_meta":{"raw":{"variants":["D3QN and knapsack team up to beat dynamic MEC baselines","Joint offload and channel scheme cuts delay-energy in moving MEC","Spatial-temporal MEC offloading solved by D3QN and knapsack","Knapsack channel allocation + D3QN offloading beats dynamic baselines","Hybrid D3QN-knapsack scheme lowers MEC delay-energy cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00121,"raw_usage":{"total_tokens":4993,"prompt_tokens":965,"completion_tokens":4028,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":3921}},"tokens_in":581,"tokens_out":4028,"duration_ms":25709,"temperature":1.0,"reasoning_tokens":3921,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:32:09.729843+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace DCA's final random assignment of specific subchannels with a channel-state-aware assignment under the same knapsack counts and rerun the simulations; if the cost drops materially, the current allocator is not optimal under the paper's own heterogeneous channel model in equation (6).","supporting_citations":[{"cited_title":"Dependency-aware computation offloading for mobile edge computing with edge-cloud cooperation,","cited_arxiv_id":null,"evidence_quote":"provides the SEG comparison baseline and simulation parameter settings"},{"cited_title":"Wireless networks for mobile edge computing: Spatial modeling and latency analysis,","cited_arxiv_id":null,"evidence_quote":"motivates the VM I/O interference model in edge-server computation time"},{"cited_title":"Joint task offloading and radio resource management in stochastic MEC systems,","cited_arxiv_id":null,"evidence_quote":"supplies the grouped-knapsack reformulation for subchannel allocation"},{"cited_title":"V-D D3QN: the variant of double deep Q-learning network with dueling architecture,","cited_arxiv_id":null,"evidence_quote":"provides the D3QN algorithm used for offloading decisions"},{"cited_title":"Deep-reinforcement-learning-based distributed computation offloading in vehicular edge computing networks,","cited_arxiv_id":null,"evidence_quote":"supplies the priority-based approach for decoupling dependent tasks"},{"cited_title":"Joint scheduling and offloading schemes for multiple interdependent compu- tation tasks in mobile edge computing,","cited_arxiv_id":null,"evidence_quote":"provides the HRROGA comparison baseline and simulation parameter settings"},{"cited_title":"Human- level control through deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"supplies the DQN mechanism with target network and experience replay that D3QN builds on"}],"review_version":1}