{"id":"0cc61fc2-8a62-44ff-a887-fec5c6ed229a","arxiv_id":"2412.18776","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A quantum-annealing-optimized QUBO formulation of virtual traffic light phase sequencing reduced simulated stopped delay and travel time versus hill climbing, simulated annealing, and actuated signals at a four-way intersection.","lead":"A team built a virtual traffic light that uses a D-Wave quantum annealer to choose which approaching connected vehicles get the right of way at an intersection. In SUMO simulations, the quantum-guided controller showed lower stopped delays and travel times than two classical heuristics and an actuated signal, but real-time latency remains above the one-second target.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported QA advantage is tested only against two stochastic heuristics on an 8-phase QUBO that is exactly solvable classically; the central claim that QA outperforms classical optimization methods is not established until an exact classical baseline is benchmarked.","rationale":"I read the paper as an empirical comparison of a D-Wave-based VTL controller against two classical heuristic controllers on a QUBO phase-sequencing model, evaluated in SUMO. The comparison is internally coherent, and the SUMO/D-Wave integration is a genuine contribution. My central concern is not the pairwise-sequential delay model in Eqs. (1) and (3), which the reader identifies as the weakest assumption, because that model is shared by all optimizers and the headline results are measured in SUMO rather than in the QUBO cost. Instead, the decisive issue is that the claim is phrased as \"quantum annealing reduces delays compared to classical optimization methods\" while the classical side contains no exact or near-optimal solver. With an 8-city AOTSP there are only 40,320 feasible permutations, and the effective decision space per call is smaller because only the first phase is executed before re-optimization. Hill Climbing and Simulated Annealing, whose hyperparameters and restart strategies are not reported, could easily be suboptimal in ways that favor the annealer. The missing exact baseline is therefore load-bearing: if an exact classical solver achieves equal or lower SUMO delays, the central comparative claim overgeneralizes. The paper already has an interactive testbed, so adding this baseline is a cheap and decisive check. I keep the reader's CONDITIONAL verdict because the work is addressable, but the condition should explicitly require benchmarking against exact classical optimization and narrowing the claim if needed.","tokens_in":17365,"tokens_out":7778,"duration_ms":72858,"concrete_test":"Run the same SUMO-in-the-loop experiment with the classical controller replaced by an exact optimizer that evaluates all feasible phase permutations (or all subset states via dynamic programming) using the Section 4 cost model and the first-phase-only execution rule, for all nine volume/zone combinations in Figs. 4 and 5 with identical traffic demand files and random seeds. If the exact classical policy yields stopped delays and travel times no worse than the D-Wave results in Tables 1-3, the headline claim that QA outperforms classical optimization methods is not supported; if exact classical is worse, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing gap is the comparison class for \"classical optimization.\" In Section 4, the VTL QUBO (Eq. 6) has only 8 phases and 64 binary variables; feasible solutions are permutations of at most 8 phases, so the optimal first phase is obtainable by exact enumeration or subset DP in well under a millisecond. Yet Section 5.3 chooses only Binary Hill Climbing and D-Wave Simulated Annealing as classical baselines, and Section 6.1.1 concludes that quantum annealing \"consistently reduces delays for CVs compared to classical optimization methods.\" Because heuristics can be arbitrarily bad depending on annealing schedules, neighborhood moves, and restarts, beating these two implementations does not establish an advantage over classical optimization. The problem is compounded by the first-phase-only execution rule in Section 4: after each solved sequence, only the first phase is used and the optimizer is called again, so the effective decision per call is even smaller than a full 8-city tour. An exact classical solver using the same Eq. (6) model is the natural missing baseline. This is a correctness-risk issue, not merely a framing issue: if exact classical control matches or beats QA delays in SUMO, then quantum annealing is not the source of the reported improvement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Virtual Traffic Light (VTL) method for an isolated 8-phase intersection, formulated as an Asymmetric Open-Loop Traveling Salesman Problem mapped to a QUBO with 64 binary variables. The QUBO is solved on a D-Wave quantum annealer, and only the first phase of the returned sequence is executed before the optimizer is called again. The authors integrate SUMO with the D-Wave cloud to evaluate the method, comparing quantum annealing (QA) against binary hill climbing and D-Wave simulated annealing, and also against a traditional actuated signal. The reported results show that QA reduces stopped delay and travel time relative to the two classical heuristics under most tested traffic volumes and VTL-zone lengths, and that QA-based VTL outperforms an actuated signal in Table 3. The paper concludes that quantum annealing significantly reduces CV delays compared with classical optimization methods.","tokens_in":17641,"tokens_out":3473,"duration_ms":32214,"significance":"If the central claim is validated, this would be a useful demonstration of quantum annealing in a traffic-control application with a credible simulation testbed. The paper's concrete contributions include the SUMO–D-Wave integration, the full 8-phase NEMA formulation, and an explicit treatment of end-to-end latency, including an honest acknowledgement that current cloud queues exceed the 1-second real-time threshold. However, the current evidence does not establish the advertised advantage over 'classical optimization methods' because the only classical comparators are two stochastic heuristics, whereas the QUBO is small enough to be solved exactly by classical enumeration in well under a millisecond. The absence of an exact classical baseline, combined with unreported simulation replication counts, leaves the central quantitative claim unsupported as stated.","major_comments":[{"comment":"The comparison class for 'classical optimization methods' is too weak. The QUBO in Eq. (6) has only 64 binary variables, and the feasible set consists of permutations of at most 8 phases, i.e., at most 8! = 40,320 feasible phase orders; moreover, only the first phase of the solved sequence is executed before re-optimization. Binary Hill Climbing and D-Wave Simulated Annealing are stochastic heuristics whose performance depends on unstated hyperparameters, so beating them does not establish an advantage over classical optimization. An exact classical solver (e.g., enumeration or subset DP on the same d_ij model) is the natural baseline and would run in far less than a millisecond. The claim in §6.1.1 that 'quantum annealing consistently reduces delays for CVs compared to classical optimization methods' is therefore not established until such an exact baseline is benchmarked.","section":"§5.3 and §6.1.1"},{"comment":"The two-sample t-tests are reported with p-values but without the number of simulation replications, the total number of vehicles per condition, or the unit of observation (per-vehicle delays vs. per-run mean delays). These details are necessary to verify the statistical claims and to interpret the non-significant results at 35% capacity against simulated annealing. The authors should report the number of independent SUMO runs per condition and the sample sizes used in each t-test.","section":"§6.1.1 and Table 1"},{"comment":"The objective function minimized by the optimizer is the sum of pairwise delays over an entire 8-phase tour, but the controller executes only the first phase of that tour and then re-optimizes. The paper does not demonstrate that minimizing the full-tour cost with Eq. (3)–(6) actually minimizes the SUMO-measured delays, or that the first phase selected by the full-tour QUBO is superior to the first phase selected by an exact one-step or short-horizon policy. A correlation analysis between optimizer cost and measured SUMO delay, or an ablation comparing full-tour re-optimization with a simpler one-step greedy rule, is needed to validate the load-bearing modeling assumption.","section":"§4 (Eqs. (1)–(6) and first-phase rule)"},{"comment":"The comparison with the traditional actuated traffic signal reports only point estimates. For 35% and 70% capacity, the stopped-delay differences are small (2.6 s and 3.6 s, respectively), and no statistical test, number of replications, variance information, or description of the Synchro timing plans is provided. The claim that QA-based VTL 'outperforms traditional traffic control systems in all traffic volumes considered' is stronger than the evidence supports without error bars or significance testing.","section":"§6.2 and Table 3"}],"minor_comments":[{"comment":"The text '264 possible choices' should read '2^64 possible choices'; the superscript formatting appears to have been lost.","section":"§4, Eq. (6) paragraph"},{"comment":"The cost-function plot lacks error bars or a description of the distribution of optimizer costs, so the statement that QA has the 'lowest average cost function value of 0.22' is difficult to interpret without knowing the scale of the d_ij values.","section":"Figure 6"},{"comment":"The ACM reference-format placeholder block in the author header is a template artifact and should be removed or replaced with the actual formatted citation.","section":"Header/ACM Reference Format"},{"comment":"There are minor typos, including 'acr oss' in §6.1.1 and the stylized 'Comparably' at the start of a paragraph in §5.3; a careful proofread would improve readability.","section":"§5.3, §6.1.1"}],"recommendation":"major_revision","confidential_remarks":"The engineering contribution—the SUMO–D-Wave quantum-in-the-loop testbed—is solid and could be a useful resource for the community. However, the central quantitative claim currently rests on an unfair comparison against two stochastic heuristics on a problem that is exactly solvable classically. Adding an exact baseline is inexpensive and should be required; if the exact baseline matches or beats QA, the paper should be reframed as a feasibility demonstration rather than a performance advantage. The statistical reporting (replication counts, test details) also needs strengthening. I would lean toward major revision rather than rejection because the identified gaps are fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new piece here is the testbed: a working SUMO-to-D-Wave loop that solves an 8-phase NEMA VTL sequencing QUBO and feeds the first phase back into the simulator. That is real engineering, and it goes beyond the earlier quantum traffic-signal papers by covering all eight NEMA movements and testing against an actuated signal as well as two classical heuristics. The authors also report the end-to-end latency honestly, including the fact that it misses the 1-second real-time target, and they do not hide the one non-significant t-test at 35% capacity. Credit where it is due: this is a plausible, clearly described demonstration.\n\nThe soft spot is the comparison class. The QUBO is an 8-city open-loop ATSP, which has only 40,320 feasible permutations. Exact enumeration or a subset DP would solve it in well under a millisecond per call, and since the optimizer is re-run after each first phase, the effective per-decision problem is even smaller. Against Binary Hill Climbing and D-Wave's Simulated Annealing, beating two stochastic heuristics does not support the paper's claim that QA outperforms classical optimization methods. This is not a framing nitpick: if an exact classical solver using the same Eq. (6) model produces the same or better SUMO delays, the quantum hardware is not the source of the improvement. The paper needs that baseline. The statistical section also omits the number of replications behind each t-test, which should be an easy fix. The delay model itself, pairing consecutive phases and ignoring platoon dispersion or spillback, is a modeling simplification; for a single isolated four-way intersection it is acceptable, though it limits how far the results generalize.\n\nWho is this for? People working on quantum optimization for transportation who want to see a concrete, working integration and a clear statement of what a VTL-QUBO looks like with all eight NEMA phases. It is not yet evidence that quantum annealing beats classical optimization for this problem class.\n\nMy recommendation: send it out. The testbed and formulation are worth referee time, and the missing exact baseline is addressable in revision. I would want the authors to add an exact classical solver, report replication counts, and soften the headline claim accordingly. That is a heavy but legitimate revision.","headline":"A real D-Wave/SUMO quantum-in-the-loop testbed for 8-phase virtual traffic lights, but the QA advantage claim needs an exact classical baseline before it means anything.","tokens_in":18195,"tokens_out":1249,"would_cite":false,"duration_ms":13648,"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":"Quantum annealing lowers stopped delay and travel time at virtual traffic lights compared with classical optimizers and actuated signals.","keywords":["virtual traffic lights","quantum annealing","QUBO","connected vehicles","traffic signal optimization","SUMO simulation","D-Wave","stopped delay"],"falsifier":"In the same SUMO scenario, replace the annealer's output with an exhaustive search over all feasible NEMA phase orders that uses the simulator's actual measured stopped delay for each order rather than the pairwise \\$d_{ij}\\$ proxy; if a phase order regularly beats the annealer's choice, the QUBO objective is not minimizing the real quantity.","tokens_in":17186,"feed_emoji":"🚦","tokens_out":7484,"duration_ms":60439,"temperature":0.7,"pith_summary":"This paper tries to establish that a quantum annealer can make virtual traffic lights (VTL) control an intersection better than classical optimization and better than a conventional actuated signal. In VTL, connected vehicles receive right-of-way based on current traffic conditions; the authors formulate the choice of which of the eight standard NEMA phases gets right-of-way, and in what order, as a QUBO problem solved on a D-Wave quantum annealer. They connect that optimizer to the SUMO microscopic traffic simulator through a cloud link and compare stopped delays and travel times under three traffic volumes and three VTL zone lengths. The reported outcome is that quantum annealing produces lower delays and travel times than hill climbing, simulated annealing, and an actuated signal in nearly all tested cases, with statistically significant t-test results except at 35% capacity versus simulated annealing. If correct, the method points toward infrastructure-free intersection control where a cloud annealer generates signal phase and timing instructions directly from connected-vehicle data.","feed_headline":"Quantum annealing cuts stopped delays at virtual traffic lights","feed_subtitle":"In SUMO tests, the D-Wave-backed VTL beat two classical solvers and an actuated signal on delay and travel time","key_machinery":"The load-bearing object is the QUBO formulation of an 8-city asymmetric open-loop TSP over NEMA phases. Each binary variable \\$x_{i,k}\\$ says whether phase \\$i\\$ occupies position \\$k\\$ in the right-of-way sequence; the cost \\$d_{ij}\\$ is the total stopped delay of vehicles in phase \\$j\\$ when phase \\$i\\$ immediately precedes it, computed from per-vehicle ETAs as \\$d_{56} = \\sum_i \\max(0, t_5 - t_{6i} + Y + R)\\$ for the example of phase 5 before phase 6. Equality constraints (every phase used once, every position filled once) become quadratic penalties with Lagrange multiplier \\$\\gamma = 100\\$, and the annealer converts the QUBO to an Ising Hamiltonian whose ground state encodes the best sequence. The mechanism that carries the argument is the claim that quantum tunneling lets the annealer escape local minima that trap hill climbing and simulated annealing.","core_discovery":"The central claim is that right-of-way sequencing at an intersection can be treated as an asymmetric open-loop traveling salesman problem over the eight NEMA phases, and that solving this sequencing problem on a quantum annealer yields lower stopped delay and travel time than solving the same QUBO with two classical heuristics or with a traditional actuated signal. The paper supports this with a quantum-in-the-loop testbed that connects SUMO to D-Wave: vehicle positions and speeds are read from the simulation, pairwise delay costs between consecutive phases are computed from estimated times of arrival, the annealer returns a phase order, and only the first phase is executed before the optimizer is called again. Across volumes of 35%, 70%, and 105% of capacity and VTL zone lengths of 50, 75, and 100 meters, the reported overall averages are 46 seconds of stopped delay for quantum annealing versus 91.1 for simulated annealing and 111.3 for hill climbing, with travel times of 115.3 versus 161.7 and 181.3 seconds. The authors also report that the method beats an actuated signal, with the largest gap at 105% capacity, where stopped delay drops from 123.4 to 53.9 seconds.","pith_inferences":["A stronger test would compare the annealer's phase order against exhaustive search over all 8! phase orders in the same simulator; the paper does not report such a check, so the annealer's advantage over classical heuristics is not yet shown to be true optimality.","Because the system executes only the first phase of each returned sequence and then re-optimizes, the multi-phase tour acts as a greedy first-phase selection; the practical benefit may hinge on the first phase rather than on the full ordering.","The evaluation assumes full connected-vehicle penetration; extending the method to mixed traffic would require estimating ETAs and delay costs for non-CV vehicles from sensors or roadside units.","The same QUBO-AOTSP framing could coordinate multiple intersections or include pedestrian phases once cloud latency shrinks, directions the authors mention but do not test."],"forward_implications":["Intersections can be controlled without fixed signal infrastructure, with the annealer generating phase and timing instructions directly from connected-vehicle data.","The same QUBO formulation drops unoccupied phases automatically, so the method adapts to lopsided demand at an intersection.","The annealer's processing time of about 197 ms per call supports real-time use; the current 2.99 s end-to-end latency is dominated by cloud upload, queue, and download.","The reported advantage over an actuated signal is largest at 105% of capacity, suggesting the biggest payoff under heavy congestion.","If quantum cloud queue times drop toward classical cloud levels, the method could meet the 1-second latency threshold for real-time mobility applications."],"supporting_citations":[{"why":"Supplies the QUBO formalism that encodes the phase-sequencing optimization.","marker":"[31]"},{"why":"Defines the standard NEMA eight-phase diagram used to enumerate intersection movements.","marker":"[51]"},{"why":"Introduces the virtual traffic light concept and system design this work builds on.","marker":"[54]"},{"why":"Earlier quantum-annealing signal control that this paper extends by adding left-turn phases.","marker":"[25]"},{"why":"Global quantum-annealing traffic signal control used as a prior approach with limited movement types.","marker":"[26]"},{"why":"Describes simulated annealing, the classical baseline that the annealer is compared against.","marker":"[43]"},{"why":"Provides evidence for quantum annealing's ability to escape local minima, the mechanism behind the claimed advantage.","marker":"[8]"},{"why":"Measures a roughly 169 ms classical cloud latency benchmark used to argue that the method becomes feasible when quantum cloud infrastructure matures.","marker":"[12]"},{"why":"Defines the open-loop traveling salesman problem whose asymmetric variant models phase sequencing.","marker":"[10]"},{"why":"Supports the statement that quantum annealers successfully optimize QUBO combinatorial problems.","marker":"[41]"}],"fun_headline_variants":["Quantum annealer beats classical solvers at virtual traffic lights","D-Wave annealer cuts intersection delays in SUMO test","Quantum-in-the-loop VTL outperforms classical optimization","Quantum annealing beats classical heuristics on intersection delay","Quantum annealer reduces travel time in virtual traffic light test"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the right-of-way plan is a single ordering of the eight phases and that the delay of putting one phase before another is fully captured by pairwise stopped-delay estimates between consecutive phases, ignoring overlapping non-conflicting movements, queue spillback, and platoon dispersion; if that premise fails, the optimizer may lower a proxy rather than real delay.","fun_headline_variants_meta":{"raw":{"variants":["Quantum annealer beats classical solvers at virtual traffic lights","D-Wave annealer cuts intersection delays in SUMO test","Quantum-in-the-loop VTL outperforms classical optimization","Quantum annealing beats classical heuristics on intersection delay","Quantum annealer reduces travel time in virtual traffic light test"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000705,"raw_usage":{"total_tokens":3255,"prompt_tokens":1100,"completion_tokens":2155,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":716,"completion_tokens_details":{"reasoning_tokens":2076}},"tokens_in":716,"tokens_out":2155,"duration_ms":14638,"temperature":1.0,"reasoning_tokens":2076,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:15:02.077770+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the same SUMO scenario, replace the annealer's output with an exhaustive search over all feasible NEMA phase orders that uses the simulator's actual measured stopped delay for each order rather than the pairwise \\$d_{ij}\\$ proxy; if a phase order regularly beats the annealer's choice, the QUBO objective is not minimizing the real quantity.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the QUBO formalism that encodes the phase-sequencing optimization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the virtual traffic light concept and system design this work builds on."},{"cited_title":"Rutenbar","cited_arxiv_id":null,"evidence_quote":"Describes simulated annealing, the classical baseline that the annealer is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the open-loop traveling salesman problem whose asymmetric variant models phase sequencing."}],"review_version":1}