{"id":"e602605d-2201-4afd-9781-dd562e408a9f","arxiv_id":"2412.15703","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A CNN-VAE global representation plus PPO yields faster training and stable control on grid networks, though the claimed superiority over baselines is inconsistent across scenarios.","lead":"This paper combines a CNN-based variational autoencoder with PPO to give traffic signal agents a compact global view of the road network, aiming for faster and more stable training than graph-based DQN methods. It also introduces a SUMO-based dynamic traffic scenario with emergency road blockages, and reports speed and stability gains.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline superiority claim may rest on poorly configured baselines: DQN baselines in Table 3 report ~2.5M waiting time versus 4.7k for MacLight, and no baseline tuning or reward details are given.","rationale":"I agree with the reader that the empirical comparison is the weakest load-bearing assumption. The specific problem is sharper than 'undertuned': the magnitude of the DQN baseline numbers in Table 3 is implausible for any published method, so the burden is on the authors to show the baselines were run in a faithful configuration. The paper provides model hyperparameters in Appendix B but no baseline hyperparameters, and the reward-selection discussion in Section 3 is conducted with IPPO on a different dataset, not with the compared methods. This is resolvable by a concrete rerun, so I do not move the verdict: conditional acceptance is right. I found no reason to doubt the internal mechanics of the CNN-VAE/PPO design, and the wall-clock advantage is plausible once the comparison is fair.","tokens_in":14279,"tokens_out":7177,"duration_ms":65153,"concrete_test":"Using the paper's released code (https://github.com/Aegis1863/MacLight) or the original public implementations, rerun CoLight, DuaLight, and IDQN on the Normal scenario with (a) their original published reward and hyperparameters and (b) the paper's waiting-time reward with a small grid search over learning rate and target-update frequency, using seeds 42–46 and 80 episodes on the same A100. If either configuration brings baseline waiting time within 2x of MacLight's 4,737 (or queue within 2x of 140), the reported superiority is an artifact of baseline configuration; if all configurations remain above ~100k waiting time, the superiority claim survives this test.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Abstract and §5.3 claim MacLight has 'superior stability, optimized convergence levels and the highest time efficiency' relative to SOTA. The evidence is Table 3 and Table 5. In Table 3, IDQN/CoLight/DuaLight report waiting times of 1.65M–2.63M in Normal, while fixed-time is 56k and MacLight is 4.7k. Such values are far outside the range of published results for these methods and suggest a mismatch in reward, state encoding, or hyperparameter configuration. Section 3 defines a waiting-time reward r=W_{t-1}-W_t, but the paper never states whether the DQN baselines were re-trained with this reward or with their original reward, nor does it report baseline hyperparameters or tuning budget. If the baselines are undertuned or forced into an unfavorable reward, the convergence and wall-clock advantages (Table 5) are not attributable to MacLight. Additionally, the paper's own ablation IPPO outperforms MacLight on every Block metric in Table 3 (e.g., return -12.0 vs -17.3), weakening the 'superior stability' claim. A rerun with the original, properly tuned baselines is the decisive check.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MacLight, a traffic signal control method that combines a CNN-based variational autoencoder for global state compression with a PPO backbone for decentralized control, and introduces a dynamic traffic scenario with random road blockages in the SUMO simulator. The method is evaluated against fixed-time, IPPO, MAPPO, IDQN, CoLight, and DuaLight on a 4x4 grid under Normal, Peak, and Block scenarios. The main reported advantages are faster wall-clock training (Table 5) and competitive or better waiting-time/queue/speed metrics in the Normal scenario, with acknowledged inferiority to IPPO in the Block scenario. The paper claims 'superior stability, optimized convergence levels and the highest time efficiency' relative to general and domain SOTA methods, but this claim is weakened by the Block-scenario result and by unresolved questions about baseline configuration and reward selection.","tokens_in":14386,"tokens_out":5349,"duration_ms":47014,"significance":"If the results are reproducible, the wall-clock training advantage in Table 5 is a practically meaningful contribution: reducing training time from several hours to under one hour on the same hardware directly addresses a known bottleneck of graph-attention DQN methods. The dynamic road-blockage scenario is also a useful extension to TSC benchmarking, and the authors provide a code link. The method itself is straightforward and its ablation against IPPO helps isolate the effect of the global representation. However, the headline superiority claims rest on a comparison whose fairness is not established, and the Block-scenario numbers contradict the stated 'superior stability' claim. The paper's contributions are defensible but require substantial additional evidence.","major_comments":[{"comment":"The DQN-based baselines (IDQN, CoLight, DuaLight) report waiting times on the order of 1.6–2.6 million in the Normal scenario, which is 29–46 times worse than the fixed-time baseline (56,409) and 350–550 times worse than MacLight (4,737). Such values are far outside the range reported in the original CoLight/DuaLight papers for comparable grid networks, and strongly suggest a mismatch in reward function, state encoding, or hyperparameter configuration. The paper never states whether these baselines were retrained with the waiting-time reward r = W_{t-1} - W_t defined in §3, nor does it report their hyperparameters, network sizes, or tuning budget. Since the central claim of superiority over domain SOTA depends on these numbers, the authors must provide baseline configuration details, run the baselines with their original pressure-based rewards, or otherwise demonstrate that the comparison is fair. This is a load-bearing issue for the headline claim.","section":"§5.3, Table 3"},{"comment":"The abstract claims MacLight achieves 'superior stability' and 'optimized convergence levels' compared to general and domain SOTA methods, but in the Block scenario the ablation IPPO outperforms MacLight on every reported metric: return -12.0 vs -17.3, waiting time 13,144 vs 24,224, queue 221 vs 249, and speed 5.4 vs 5.2. The text acknowledges 'indicators are inferior to IPPO' in §5.3, yet the abstract and conclusion still assert general superiority. The paper attempts to explain this by saying that average metrics over the whole process hide the final performance, and Table 4 shows near-tie test returns (-1.18 vs -1.17). This is not sufficient: the training-phase metrics are the advertised results, and the dynamic Block scenario is a stated novelty of the paper. The authors should either temper the 'superior stability' claim to exclude Block, provide a mechanistic explanation for why the global representation degrades the dynamic scenario, or modify the method to recover the Block performance.","section":"§5.3, Table 3"},{"comment":"The statistical support for 'superior stability' is weak. With five seeds, the standard deviations are very large: in Peak, MacLight's return is -362.3 ± 423 while IPPO is -434.8 ± 451, and in Normal the returns are -4.02 ± 10 versus -6.6 ± 22. No significance tests or confidence intervals are reported. The visual stability claim from Fig. 7 is also not quantified. The authors should report per-seed results and statistical tests (e.g., matched-pair comparisons across the same seeds) to substantiate the stability claim, especially where the mean differences are smaller than the standard deviations.","section":"§5.1, Table 3"},{"comment":"The reward function is selected based on experiments in the 'ingolstadt21' scenario, which the authors state is 'completely different' from their own 4x4 grid, and the selection is made using only IPPO. This leaves open the possibility that the chosen waiting-time reward is specifically favorable to PPO-style methods and unfavorable to DQN-based baselines. The paper does not report reward ablation on the actual Normal/Peak/Block scenarios, nor does it compare against the pressure reward used by CoLight/DuaLight. Given that the baseline comparison is the basis for the SOTA claim, the reward choice is a free parameter that should be justified on the actual test scenarios, or the authors should show that the baselines perform comparably under their original reward.","section":"§3, Table 1"}],"minor_comments":[{"comment":"The text says 'We first give the process of upsampling' but Eq. (1) describes the encoder (downsampling) path. The wording should be corrected to 'downsampling'.","section":"Eq. (4)"},{"comment":"The notation 'ConvTrans33_3' is ambiguous; it likely means a transposed convolution with 33 output channels and kernel size 3, but the subscript/superscript convention is inconsistent with Eq. (1). Please define the notation unambiguously.","section":"Abstract"},{"comment":"There is a typo: 'domian' should be 'domain'.","section":"§4.1 and §6"},{"comment":"The grid-based limitation is acknowledged in the conclusion but not in the abstract. Since the abstract claims general 'multi-scene aggregation' without mentioning the regular-grid assumption, the authors should state this limitation more prominently, as it restricts the method to Manhattan-like networks.","section":"Table 3"},{"comment":"The caption says 'The specific values in the table include the mean of the current column indicator and the standard deviation in brackets,' but the Waiting Time columns in Table 3 do not show standard deviations. Please either include them or explain why they are omitted.","section":"§3"},{"comment":"The local observation vector is said to have four components (one-hot action, switching boolean, vehicle density, waiting density), but the total dimension is later given as 33 in Appendix B. The mapping from these components to 33 is not explained; please provide the exact vector construction.","section":"Appendix B"},{"comment":"In the ValueNet parameters, state_dim is listed as 33 and global_emb_dim as 16. If the value input is the concatenation of local and global representations, the input dimension should be 49. Please clarify whether state_dim refers only to the local part or whether the concatenation is handled elsewhere.","section":"§4.4"},{"comment":"The dynamic scenario description does not specify the duration of the road blockage, the rerouting mechanism (e.g., SUMO rerouter settings), or how many episodes experience the blockage. These details are necessary for reproducibility and for interpreting the Block-scenario results.","section":"§1"},{"comment":"The claim that 'none of the current peer studies have deployed dynamic traffic systems for experiments' is stated without a survey citation to support it. Several works have used varying demand patterns or incidents; please soften the claim or provide evidence that dynamic emergency scenarios have not been tested before.","section":"§5.3"},{"comment":"The statement 'if we check the final value, MacLight still has the best performance' in the Peak scenario is not supported by a table or figure that reports final-value metrics; Table 4 only reports test returns, not waiting time/queue/speed at the final episode. Please add explicitly reported final metrics.","section":"§5.3"}],"recommendation":"major_revision","confidential_remarks":"The main risk to the paper's acceptance is the fairness of the baseline comparison. The reported waiting times for IDQN/CoLight/DuaLight are so far outside their published performance that either the reward function, the state encoding, or the hyperparameters must be mismatched. I recommend asking the authors to provide baseline training details (hyperparameters, tuning budget, reward used) and, if necessary, to rerun the DQN baselines with their original pressure-based reward. The Block-scenario result where the ablation IPPO beats the full MacLight method also needs to be addressed head-on rather than acknowledged in passing. The current manuscript's abstract overstates the findings relative to Table 3 and Table 5. These issues are fixable with additional experiments and revised claims, so major revision is appropriate rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: MacLight is a workmanlike empirical TSC paper with one genuinely useful idea—using an online CNN-VAE global embedding to condition the PPO value module—and a new dynamic blockage scenario. The wall-clock advantage in Table 5 is credible: 40-58 minutes per 80 episodes versus hours for the DQN-based baselines is a real difference if those baselines are implemented fairly. The ablation design (IPPO vs MacLight) is also a plus.\n\nThe headline overclaims, though. The abstract says MacLight achieves \"superior stability\" and \"optimized convergence levels\" relative to SOTA, yet in the Block scenario IPPO beats MacLight on every metric (return -12.0 vs -17.3, waiting time 13.1k vs 24.2k). The paper admits this in Section 5.3, so the abstract is not consistent with its own data. The comparison with CoLight, DuaLight, and IDQN is compromised: the paper never states whether those baselines were re-trained with the same waiting-time reward or kept their original pressure-based rewards, and it reports no baseline hyperparameters or tuning budget. The waiting times in Table 3 (1.6M-2.6M) are orders of magnitude worse than fixed-time (56k), which suggests the baselines are either undertuned or mismatched to the reward. That does not invalidate the wall-clock speed claim—that is about implementation—but it does invalidate any \"superiority\" claim over SOTA in return or stability.\n\nOther soft spots: the reward function was selected by testing on ingolstadt21, a different dataset, then applied to all algorithms without checking whether it suits the DQN baselines; standard deviations are large (e.g., IPPO return -6.6 ± 22); and no significance tests are reported. The \"first dynamic traffic\" claim also needs a literature search—the paper does not demonstrate that no prior work used dynamic incidents in SUMO.\n\nIf the authors rerun the DQN baselines with their original, properly tuned rewards and add per-scenario significance tests, the core method and timing results would be solid. As is, this is a plausible incremental contribution with an overreaching abstract. For a TSC researcher, it is worth a skim for the dynamic scenario and the CNN-VAE idea. I would send it to review if it were a fresh submission; the flaws are fixable. But I would not cite it as evidence of SOTA superiority yet.","headline":"Plausible speed win and a useful new dynamic-blockage scenario, but the abstract overclaims stability and the baseline comparison is compromised by missing tuning/reward details.","tokens_in":15058,"tokens_out":2310,"would_cite":false,"duration_ms":21730,"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":"Traffic-light agents train under an hour via scene compression","keywords":["Traffic signal control","Multi-scene convolution","Variational autoencoder","Multi-agent reinforcement learning","PPO","Dynamic traffic simulation","SUMO","Global state representation"],"falsifier":"Re-run Normal, Peak, and Block with CoLight, DuaLight, and IDQN using their originally reported pressure-based rewards and tuned hyperparameters on the same SUMO setup, and check whether MacLight's wall-clock and final waiting-time advantages persist; alternatively, test MacLight on a non-grid road network to see whether the CNN aggregation matrix assumption breaks.","tokens_in":13955,"feed_emoji":"🚦","tokens_out":8253,"duration_ms":67044,"temperature":0.7,"pith_summary":"MacLight aims to show that traffic-signal control becomes faster and more stable when global road information is compressed by a convolutional variational autoencoder and fed to the value network of a PPO controller, rather than modeled as a graph-attention deep Q-network. The paper tests the method on a 4x4 SUMO grid under normal, heavy, and dynamically blocked traffic, reporting training times of 39 to 58 minutes per 80 episodes against 137 to 456 minutes for DQN-based baselines. It also introduces a dynamic-traffic benchmark in which random roads close and vehicles reroute, which the authors say previous peer studies did not deploy. The practical interest is that training cost and policy collapse are the main barriers to putting learned signal control in the field, and the paper attacks both with a single architecture.","feed_headline":"Traffic-light agents train under an hour via scene compression","feed_subtitle":"A convolutional autoencoder feeds a global view to the PPO critic, beating DQN graph methods on speed and stability.","key_machinery":"The load-bearing object is the multi-scene aggregation matrix: each intersection's local feature vector is oriented and placed into its geographic slot in a tensor whose width and height are grid positions and whose channels are feature values. A CNN-based variational autoencoder compresses this tensor into a compact latent representation, and the PPO value function consumes the concatenation of that representation with the local observation while the policy network improves from local features alone, letting global context guide value estimation without graph attention or neighbor sampling.","core_discovery":"The central discovery is that separating the global scene representation from the local intersection view, and feeding the global code only into the PPO value function while the policy keeps local features, prevents the overfitting and policy collapse seen in DQN-based graph methods. A CNN-based variational autoencoder is trained online to compress the scene matrix into a compact latent code; the value network evaluates the concatenation of that code with the local observation, and training stays stable because PPO clips its updates. On the 4x4 grid, MacLight produces the best average waiting time and queue length in the normal scenario, competitive returns under peak load and road blockages, and wall-clock training of 39 to 58 minutes for 80 episodes, while IDQN needs 137 to 186 minutes and CoLight and DuaLight need 283 to 456 minutes. The authors attribute the speed to convolution's GPU parallelism and the stability to on-policy PPO updates, and they position the dynamic blockage environment as a first step toward emergency traffic scenarios.","pith_inferences":["The paper does not ablate the VAE against a cheaper compressor, so a reader could test whether simple pooling of the scene matrix preserves the speed and stability gains; that would isolate whether the latent compression or the PPO backbone is the actual driver.","All algorithms in the comparison use the waiting-time reward, while CoLight and DuaLight were originally published with pressure-based rewards; re-running those baselines with their native rewards is the natural check on whether the reported margin is a property of MacLight or of the reward choice.","The grid assumption limits immediate transfer to non-Manhattan road networks; a multiscale convolution with imputed missing neighbors is a testable extension the authors themselves point toward."],"forward_implications":["Training times below one hour per scenario on a single A100 make repeated tuning and large-scale grid experiments affordable with the same hardware.","On-policy PPO updates avoid the policy collapse that the paper reports for IDQN, CoLight, and DuaLight under sparse rewards, opening learned signal control to high-pressure and emergency conditions.","Because global context enters only the value evaluation module, the architecture keeps the parallel-friendly convolutional computation and avoids the sequential neighbor aggregation of graph models.","The random road-blockage and rerouting procedure gives a concrete stress test for whether a traffic-signal policy can absorb sudden changes in traffic distribution."],"supporting_citations":[{"why":"Supplies the PPO backbone, including the clipped surrogate objective and GAE advantage estimation that MacLight builds on.","marker":"[19]"},{"why":"Provides the variational autoencoder training objective used to compress the global scene matrix into a latent code.","marker":"[10]"},{"why":"CoLight is the graph-attention DQN baseline that MacLight compares against, and its one-hop-neighbor finding motivates the move to global aggregation.","marker":"[29]"},{"why":"DuaLight is the domain SOTA baseline built on CoLight that MacLight must outperform.","marker":"[15]"},{"why":"Defines the local observation components and the waiting-time reward that the paper adopts for all agents.","marker":"[2]"},{"why":"SUMO provides the simulation platform, including the vehicle rerouting that creates the dynamic blockage scenario.","marker":"[13]"},{"why":"Defines the independent PPO (IPPO) baseline that also serves as the ablation for MacLight's global value input.","marker":"[5]"},{"why":"Defines the MAPPO baseline whose global value network is contrasted with MacLight's local-global value design.","marker":"[31]"},{"why":"The ingolstadt21 benchmark is used to compare reward functions and justify the waiting-time reward adopted in the experiments.","marker":"[3]"}],"fun_headline_variants":["Under an hour: traffic-light training via scene compression","Scene compression speeds traffic signal training 5x","PPO and autoencoder keep traffic control stable, fast","MacLight: stable traffic signals in under an hour"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The speed-and-stability claim rests on the comparison being fair to the older DQN methods and on a regular 4x4 grid standing in for real road networks; if either gives way, the reported advantage could shrink.","fun_headline_variants_meta":{"raw":{"variants":["Under an hour: traffic-light training via scene compression","Scene compression speeds traffic signal training 5x","PPO and autoencoder keep traffic control stable, fast","MacLight: stable traffic signals in under an hour"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000347,"raw_usage":{"total_tokens":1923,"prompt_tokens":988,"completion_tokens":935,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":872}},"tokens_in":604,"tokens_out":935,"duration_ms":8135,"temperature":1.0,"reasoning_tokens":872,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:09:31.648916+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run Normal, Peak, and Block with CoLight, DuaLight, and IDQN using their originally reported pressure-based rewards and tuned hyperparameters on the same SUMO setup, and check whether MacLight's wall-clock and final waiting-time advantages persist; alternatively, test MacLight on a non-grid road network to see whether the CNN aggregation matrix assumption breaks.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DuaLight is the domain SOTA baseline built on CoLight that MacLight must outperform."},{"cited_title":"Alegre, Ana L","cited_arxiv_id":null,"evidence_quote":"Defines the local observation components and the waiting-time reward that the paper adopts for all agents."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SUMO provides the simulation platform, including the vehicle rerouting that creates the dynamic blockage scenario."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the MAPPO baseline whose global value network is contrasted with MacLight's local-global value design."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The ingolstadt21 benchmark is used to compare reward functions and justify the waiting-time reward adopted in the experiments."}],"review_version":1}