{"id":"b05ada59-fee2-4794-a66f-f8a9cb9c6c95","arxiv_id":"2504.18031","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A joint MAB-MCTS algorithm for eVTOL task offloading and trajectory planning in cognitive radio networks is proposed and evaluated in simulation.","lead":"This paper combines a Multi-Armed Bandit with Monte Carlo Tree Search to pick which ground base stations an electric air taxi should visit for task offloading, while managing battery use and spectrum conflicts. It is a simulation study claiming better task completion and energy efficiency than two simpler benchmarks, but the implementation and evidence have serious gaps.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 2, which produces the reported results, is not MCTS: it is two nested UCB loops with no expansion, rollout, or backpropagation, so the paper's central MCTS-based joint-optimization claim is unsubstantiated as written.","rationale":"I read the paper's central claim as: a MAB-MCTS framework jointly optimizes eVTOL trajectory and offloading, and the simulations validate that framework. The condition that must hold is that the method labeled MCTS in the evaluation actually is MCTS. Algorithm 2 is the executable specification, and it contains no tree expansion, no simulation/rollout, and no backpropagation; it is a nested UCB argmax over phases and BSs with failure-triggered re-selection. This is an internal inconsistency rather than a disagreement with any external consensus, and it is load-bearing because every numerical claim, including the regret comparisons in Figure 8, is attributed to the MCTS machinery. The surrogate-reward concern in Eq. (10) is real but secondary: even a correct surrogate would not fix the mismatch between the described and implemented algorithm. I therefore agree with the reader's REJECT verdict, but the reader's listed weakest assumption (unspecified weights in Eq. (10)) is not the same as the most load-bearing point, which is the algorithm identity itself. The concrete check is a traced execution of Algorithm 2 followed by rerunning the comparisons with a true MCTS implementation; absent that check, the central claim is unsupported.","tokens_in":19646,"tokens_out":6308,"duration_ms":55759,"concrete_test":"Reproduce the N=5 scenario of Section V with an execution tracer around Algorithm 2. Log every call to tree expansion, rollout/simulation, and backpropagation; if, as the pseudocode implies, zero such calls occur, the evaluation did not test MCTS. Then replace the nested UCB loops with a genuine MCTS/UCT implementation (for example Kocsis-Szepesvari UCT) and rerun the comparisons behind Figures 5 and 6. If the reported 60% energy-efficiency gain and 17% task-completion gain do not persist, the central MCTS-based claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MAB-MCTS jointly optimizes eVTOL trajectory and offloading decisions, and that this explains the reported gains. The load-bearing premise is that the algorithm evaluated in Section V is the proposed MCTS algorithm. That premise is internally contradicted by Algorithm 2 (Section IV-B): the pseudocode computes UCB values for time phases and BSs, selects the argmax of each, and on failure removes that choice and repeats. It never expands a node, never performs a rollout/simulation, and never backpropagates a value. The four MCTS steps described in the text (selection, expansion, simulation, backpropagation) are absent from the executable specification. Consequently Figures 5-7 and Tables III-IV, presented as validating the proposed MCTS approach, in fact validate an iterative UCB scheduler under a different name. Figure 8 claims backpropagation explains the regret gap between UCT and the proposed method, but Algorithm 2 has no mechanism by which a rollout updates any tree value. A related weakness is Eq. (10): the weighted surrogate reward has no proof connecting it to the MINLP objective P1, and weights c1, c2, c3 are unspecified. Both problems point to the same gap: the paper neither implements the MCTS it describes nor proves optimality of the UCB rule it actually runs. The 60% and 17% improvement claims must be re-established for a genuine MCTS procedure before the central claim can be accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a joint trajectory and task-offloading optimization framework for eVTOL swarms in cognitive radio networks. The authors formulate a mixed-integer nonlinear program (P1) that minimizes a weighted combination of energy consumption and the negative task-completion probability, subject to time, kinematic, and battery constraints. The proposed solution has two stages: a Multi-Armed Bandit (MAB) model with UCB-based selection to estimate time-varying BS spectrum availability, and a Monte Carlo Tree Search (MCTS) algorithm to choose BSs and access time windows, with adaptive re-planning on access failure. Numerical comparisons against TSP and epsilon-greedy baselines report up to 60% energy-consumption improvement and 17% higher task-completion probability, and the paper includes additional regret comparisons and a robustness check under a Poisson resource distribution.","tokens_in":19821,"tokens_out":6954,"duration_ms":70975,"significance":"The problem addressed is timely and practically relevant: eVTOL task offloading under dynamic BS spectrum and CPU availability, with energy and time constraints, is an important AAM challenge. The re-planning mechanism for access failures is well motivated, and the inclusion of a Poisson-distribution robustness check is a useful attempt to go beyond the Gaussian assumption. If the MCTS-based joint optimization were actually implemented and validated as described, the reported gains would be significant. However, the manuscript in its current form does not deliver an executable MCTS algorithm, does not justify the surrogate reward used by the decision rule, and does not provide enough reproducibility detail (no code, no values for weighting coefficients or Gaussian parameters, no trial counts) to substantiate the numerical claims.","major_comments":[{"comment":"Algorithm 2 does not implement the Monte Carlo Tree Search described in the prose. Lines 5-18 of the pseudocode compute UCB values for time phases and BSs, select the argmax of each, and on failure deselect and repeat; there is no expansion of a tree node, no simulation or rollout to a terminal state, and no backpropagation of a value along a visited path. The four MCTS steps listed in Section IV-B (selection, expansion, simulation, backpropagation) are absent from the executable specification, and Example 1's node counts are not produced by any equation in the paper. Consequently, Figures 5-8 and Tables III-IV validate an iterative UCB scheduler, not the proposed MCTS method, and the claim in Section V-E that backpropagation explains the regret gap between UCT and the proposed method has no algorithmic basis.","section":"Section IV-B, Algorithm 2"},{"comment":"The surrogate reward in Eq. (10) is introduced without any derivation connecting it to P1. Since θ'(a) is maximized, the term +c3 W_i rewards higher energy consumption unless c3 is negative, yet no sign constraints on c1, c2, c3 are given anywhere, and no values or sensitivity analysis are reported. Moreover, the Remark following Definition 1 states that the mean CPU cycle duration L_cp can only be obtained when the swarm is already connected to a BS, which makes it infeasible to evaluate \\L_cp(t_d) in Eq. (10) for unvisited BSs during decision-making. Step (a) in Eq. (11) also equates hovering time to \\L_cp(t_d) and omits the communication time T_com, contradicting constraint C2. As a result, the paper provides no valid argument that maximizing Eq. (10), with any choice of weights, solves or approximately solves the stated MINLP P1.","section":"Section IV-B, Lemma 4, Eq. (10)-(11)"},{"comment":"Algorithm 1 does not match the UCB solution of Lemma 3. The pseudocode uses an epsilon-greedy style comparison (\"if random number ≤ ε\"), refers to the undefined \"parameters of greedy algorithm\", and only loosely says to \"Calculate each station's UCB value\", without ever computing the index in Eq. (7). It also includes the unexplained step \"Formulate a tree structure\". Since Algorithm 1 is the stated source of the estimated spectrum availability \\P_sp(t_d) used by the subsequent decision-making stage, this mismatch undermines the MAB estimation component of the proposed framework.","section":"Section IV-A, Algorithm 1"}],"minor_comments":[{"comment":"The units are inconsistent: Table II lists the required CPU cycle duration in seconds (20-40), while Section V states 20 minutes, 30 minutes, and 40 minutes for the 5-BS, 7-BS, and 10-BS cases, respectively; Fig. 6 y-axis label \"Energy consumption (w)\" also mixes power units with energy quantities.","section":"Section V, Table II and text"},{"comment":"The caption of Fig. 8(c) repeats \"N=5 BSs\" instead of \"N=10 BSs\", and the caption of Fig. 9(c) has the same repetition; these should be corrected.","section":"Figures 8 and 9 captions"},{"comment":"The sign convention for the weighting coefficients is never stated; even if negative c3 is intended, its value and interpretation should be given explicitly because Eq. (10) is central to all numerical experiments.","section":"Section IV-B, Lemma 4"},{"comment":"The Remark after Definition 1 says CPU-cycle information is not available during pre-learning, but Eq. (10) requires \\L_cp(t_d) at decision time; the paper needs to explain how this quantity is obtained or replace it with a learnable prior.","section":"Section III-B and Section IV-B"},{"comment":"No statistical details are reported: there are no error bars, confidence intervals, or numbers of Monte Carlo trials, which is particularly important because some reported differences, such as the N=10 Scenarios 3-4 case, are small.","section":"Section V-A, Fig. 5"},{"comment":"The comparison against \"UCT\" is not reproducible because no pseudocode or implementation details for UCT are given, and the claim that UCT suffers nearly 100% higher regret due to a lack of backpropagation cannot be checked from the manuscript.","section":"Section V-E, Fig. 8"},{"comment":"Constraints C4 and C5 reference q[t] and v[t], but these variables are not defined in the optimization statement; the weight λ in the objective is also never assigned a value in the simulations.","section":"Section III-C, P1"},{"comment":"The symbol T is used both for the mission deadline in P1 and for the horizon in the MAB regret expression in Eq. (3), while T_d is listed as the number of discrete time periods and t_d is the period index; this overloaded notation should be disambiguated.","section":"Table I and Section III"}],"recommendation":"reject","confidential_remarks":"The core problem is that the executable algorithms in the paper do not implement the methods described in the text: Algorithm 2 is not MCTS, and Algorithm 1 is not UCB. The numerical results therefore do not validate the paper's central claim. Repairing this would require re-implementing the algorithmic core and re-running all experiments, which is effectively a new manuscript rather than a routine revision. I also note that the surrogate reward appears to be incomputable as stated because CPU-cycle information is declared unavailable before connection, and no parameter values are reported for the weighting coefficients."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The problem is real, but the paper's central claim is not supported by what is actually implemented. Algorithm 2, the thing that produces the reported results, is not a Monte Carlo tree search. It is two nested UCB loops: pick the best time phase by UCB, then pick the best BS by UCB, and repeat after failures. There is no node expansion, no rollout, no backpropagation. The four MCTS phases described in Section IV-B never appear in the pseudocode. So the 60%/17% gains and the regret plots validate an iterative UCB scheduler, not the proposed MCTS. This is a load-bearing mismatch, not a typo.\n\nWhat the paper does well: the system model is coherent. Formulating the joint trajectory, BS selection, and offloading time problem for eVTOL swarms under time-varying CR resources is a legitimate extension of the UAV-offloading literature. The MAB-based pre-learning of spectrum availability is a sensible idea, and comparing against TSP and epsilon-greedy is a reasonable first benchmark set.\n\nThe soft spots go beyond the algorithm mismatch. Equation (10) introduces a weighted surrogate reward with no derivation linking it to the MINLP objective P1; the weights c1, c2, c3 are unspecified. The simulation section lacks error bars, gives no energy or velocity parameters, and the text conflicts with the table on time units (40-70 seconds in Table II versus '20 minutes' in the text). The abstract's claim of up to 60% energy improvement is not clearly supported by the aggregate figures. Figure 8 attributes the regret gap to backpropagation, but Algorithm 2 has no such mechanism.\n\nWho is this for? Readers working on eVTOL offloading in cognitive radio networks might find the problem formulation and the UCB heuristic useful as a starting point. But as a validation of MCTS-based joint optimization, the paper fails. The authors need to either implement real MCTS or relabel Algorithm 2 as an iterative UCB scheduler, add reproducibility details, and re-evaluate. I would not send this to peer review in its current form; it would be a waste of reviewer time. A competent referee would hit the same wall within minutes. The underlying problem is worth another shot after a major rewrite.","headline":"Real problem, sensible formulation, but Algorithm 2 is not MCTS and the reported gains are not what they seem.","tokens_in":20495,"tokens_out":3404,"would_cite":false,"duration_ms":32828,"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":"Tree-search planner lifts eVTOL task success by 17 percent","keywords":["eVTOL","Advanced Air Mobility","cognitive radio","Monte Carlo Tree Search","multi-armed bandit","trajectory optimization","task offloading","spectrum availability estimation"],"falsifier":"Take a small instance—five base stations and three time periods—enumerate every feasible BS sequence and time-slot assignment, and compare the true objective P1 value of the exhaustive optimum with the trajectory produced by the MCTS planner under the paper's default weights; if the MCTS trajectory's true task-completion probability falls below the exhaustive optimum by more than the reported 17 percent margin, the surrogate reward is misaligned with the stated objective.","tokens_in":19309,"feed_emoji":"🚁","tokens_out":7059,"duration_ms":63184,"temperature":0.7,"pith_summary":"Electric vertical take-off and landing (eVTOL) aircraft must offload safety-critical computation to ground base stations whose spectrum and CPU resources shift with the time of day, and this paper tries to plan the offloading so that tasks finish within time and battery limits. The central claim is that the joint choice of which base stations to visit, in what order, and for how long can be solved by separating temporal resource estimation from spatial route search. A multi-armed bandit learns each station's spectrum-availability pattern first; a Monte Carlo Tree Search then scores candidate station–time pairs by a weighted mix of spectrum availability, expected CPU cycle length, and flight energy. Simulations against shortest-path and epsilon-greedy baselines report up to 60 percent energy-efficiency gains and a 17 percent higher task-completion probability in resource-limited settings. A sympathetic reader would come away with a two-stage planner that also re-plans mid-flight when a station fails to deliver its promised resources.","feed_headline":"Tree-search planner lifts eVTOL task success by 17 percent","feed_subtitle":"Bandit-learned spectrum and CPU availability feed a tree search that picks base stations and routes within battery limits.","key_machinery":"The load-bearing object is the modified UCB reward of Eq. (10), $\\theta'(a) = c_1 \\hat{P}_{\\mathrm{sp}}(t_d^*) + c_2 \\hat{L}_{\\mathrm{cp}}(t_d^*) + c_3 W_i$, which converts the three competing concerns — spectrum availability, CPU-cycle length, and interval energy cost — into a single score used to select both time periods and base stations inside the MCTS tree. The tree itself alternates time-period nodes and base-station nodes, with backpropagation updating empirical success counts (as in Example 1), and access failures trigger re-expansion of the tree. The paper's complexity analysis gives the search a per-iteration cost of $O(n \\cdot (b \\log_b t + k))$, with the bandit prior argued to shrink the effective search space.","core_discovery":"The paper's core discovery is that the mixed-integer nonlinear program P1 — minimizing flight-plus-hover energy while maximizing offloading success probability under time, velocity, acceleration, and battery constraints — can be effectively attacked by a bandit-plus-tree-search pipeline without convex relaxation or exhaustive search. The bandit stage estimates the probability that each base station has an idle spectrum hole at each time period, using an upper confidence bound rule, and records the mean available CPU-cycle duration when a connection is made. The tree-search stage builds a decision tree whose nodes alternate time periods and base stations; each node's value is the modified UCB reward in Eq. (10), a weighted sum of estimated spectrum availability, estimated CPU processing time, and the energy cost of the travel interval. When execution reveals a failed access, the failed station is removed and the tree is re-expanded, so the same machinery handles both offline pre-planning and online re-planning. The reported 17 percent task-completion gain and 60 percent energy-efficiency gain are the claimed evidence that this two-stage architecture outperforms visiting all stations via the shortest path and epsilon-greedy exploration.","pith_inferences":["Because Eq. (10) leaves the weights $c_1, c_2, c_3$ unspecified, a natural extension is to tune them against a small set of validated scenarios or derive them from the KKT conditions of a Lagrangian relaxation of P1; the paper gives no procedure for setting them.","Since the bandit and the tree search share the same UCB statistic, the framework extends naturally to multiple competing swarms by treating other swarms as additional primary users, an extension the paper names as future work.","The 60 percent energy-efficiency figure is a simulation outcome for specific counts of base stations; a hardware-in-the-loop test on a small eVTOL testbed comparing measured battery draw and completed offloads would tell whether the gain survives real propulsion and communication dynamics.","The re-planning on failure behavior is effectively online planning under model misspecification; bounding the regret of the MCTS planner against a clairvoyant offline optimizer with biased resource estimates would give the method a firmer theoretical footing than the reported simulations."],"forward_implications":["If the algorithm works as claimed, eVTOL operators can pre-plan an offload route before takeoff and re-plan it in flight without needing an exhaustive optimizer, because the MCTS search stays tractable as the number of base stations grows.","In resource-scarce networks the method's advantage appears as higher task-completion probability (up to 17 percent); in resource-rich networks the advantage shifts to lower energy use, since it visits fewer base stations than exhaustive routing.","The MAB estimation stage means the planner does not depend on a specific spectrum-availability distribution; the paper reports consistent task-completion behavior when availability follows a Poisson rather than a Gaussian law.","In denser networks the gap over the epsilon-greedy baseline widens, suggesting the method scales better than fixed-random exploration as the search space grows."],"supporting_citations":[{"why":"Supplies the UCB confidence-bound formula (Eq. 7) that the MAB estimator uses.","marker":"[38]"},{"why":"Provides the hovering and propulsion power model (Eq. 2) used in the energy constraint.","marker":"[37]"},{"why":"Serves as the UCT baseline in the accumulated-regret comparison (Fig. 8).","marker":"[39]"},{"why":"Justifies the Gaussian model for spectrum-availability probabilities used in the system model.","marker":"[36]"},{"why":"Exemplifies prior task-offloading work that ignores primary-user access conflicts, the gap this paper targets.","marker":"[11]"}],"fun_headline_variants":["Bandit-guided tree search boosts eVTOL offloading by 17%","MCTS plus bandits: 17% better eVTOL task offloading","Tree search plans eVTOL routes to lift task success 17%","eVTOL trajectory optimizer: 17% success, 60% energy savings","Smart tree search hits 17% gain for eVTOL offloading"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline rests on the surrogate reward in Eq. (10), whose three weights are set without a derivation linking them to the original MINLP objective; if those weights are chosen poorly, the route the tree search returns is not optimal for P1.","fun_headline_variants_meta":{"raw":{"variants":["Bandit-guided tree search boosts eVTOL offloading by 17%","MCTS plus bandits: 17% better eVTOL task offloading","Tree search plans eVTOL routes to lift task success 17%","eVTOL trajectory optimizer: 17% success, 60% energy savings","Smart tree search hits 17% gain for eVTOL offloading"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000428,"raw_usage":{"total_tokens":2245,"prompt_tokens":1056,"completion_tokens":1189,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":672,"completion_tokens_details":{"reasoning_tokens":1088}},"tokens_in":672,"tokens_out":1189,"duration_ms":10252,"temperature":1.0,"reasoning_tokens":1088,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:26:21.968806+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small instance—five base stations and three time periods—enumerate every feasible BS sequence and time-slot assignment, and compare the true objective P1 value of the exhaustive optimum with the trajectory produced by the MCTS planner under the paper's default weights; if the MCTS trajectory's true task-completion probability falls below the exhaustive optimum by more than the reported 17 percent margin, the surrogate reward is misaligned with the stated objective.","supporting_citations":[{"cited_title":"Sample-efficient neural architecture search by learning actions for monte carlo tree search,","cited_arxiv_id":null,"evidence_quote":"Supplies the UCB confidence-bound formula (Eq. 7) that the MAB estimator uses."},{"cited_title":"Uav-enabled wireless power transfer with base station charging and uav power consumption,","cited_arxiv_id":null,"evidence_quote":"Provides the hovering and propulsion power model (Eq. 2) used in the energy constraint."},{"cited_title":"Bandit based monte-carlo planning,","cited_arxiv_id":null,"evidence_quote":"Serves as the UCT baseline in the accumulated-regret comparison (Fig. 8)."},{"cited_title":"Analytical evaluation of outage probabil- ity based on signal to interference ratio for gaussian-distributed users,","cited_arxiv_id":null,"evidence_quote":"Justifies the Gaussian model for spectrum-availability probabilities used in the system model."},{"cited_title":"Intelligent task offloading for heterogeneous v2x communications,","cited_arxiv_id":null,"evidence_quote":"Exemplifies prior task-offloading work that ignores primary-user access conflicts, the gap this paper targets."}],"review_version":1}