{"id":"c05fd2b7-27ec-4871-b7b2-af1d005bfc7c","arxiv_id":"1908.10823","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An evolving finite-state machine using online clustering and action-conditioned Markov chains recognizes a collision-prone Dead-End state and predicts subsequent states in a simulated car-following task.","lead":"An online 'evolving Finite State Machine' clusters driving situations as they happen and learns a separate Markov transition matrix for each control action. The paper shows the method can recognize a collision-prone Dead-End state and predict future states in a simple simulated car-following scenario.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim is not reproducible: Eq. 7 uses an undefined cluster-center variance var(z*_i), so the uniqueness of Dead-End state #3 and the JSD validation rest on an uncomputable quantity.","rationale":"The reader's weakest_assumption already identifies the undefined cluster variance in Eq. 7 as a load-bearing issue; my stress-test confirms this is the most critical point, because every experimental result in Sections 3.2.1 and 3.2.2 flows through that equation. The undefined variance makes the method non-reproducible and the claimed uniqueness of Dead-End state #3 an unverifiable implementation detail. I also note the likely determinism of the repeated runs, which further weakens the 'consistently recognized' phrasing, but the undefined variance is the more fundamental flaw. Since the paper can be repaired by defining the quantity and re-running with stochastic seeds, the appropriate verdict remains CONDITIONAL, as the reader originally concluded. No change in verdict is required, but the revision must address the Eq. 7 definition and add truly independent trials.","tokens_in":10749,"tokens_out":3807,"duration_ms":41530,"concrete_test":"Provide the missing definition of var(z*_i) in Eq. 7 (e.g., empirical variance of observations assigned to cluster i, or a fixed positive scale) and re-run all four cases of Section 3.2 using 20 repetitions with small random perturbations to the initial headway and vehicle speeds. If any collision maps to a state other than the same unique index, or if the held-out JSD exceeds 0.15, the uniqueness and consistency claims fail.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the Dead-End state is 'uniquely determined and consistently recognized' (Section 3.2.1) and that transition matrices are 'precisely identified' (Section 3.2.2)—depends on the state-recognition probabilities from Eq. 7. That equation contains the term var(z*_i), the variance of a cluster center, but no definition of this quantity is provided anywhere in the manuscript. A cluster center is a single vector, so its variance is not well-defined without an explicit convention (e.g., variance of the observations assigned to the cluster, or a user-specified scalar). Since Prob(St), the transition-matrix updates (Eqs. 15–17), the predicted distributions Prob_pred(St+1), and the JSD results in Figure 5 are all computed from Eq. 7, the experimental evidence cannot be reproduced or independently checked. The mapping of all collision cases to 'state #3' is therefore an artefact of the specific, unreported implementation of Eq. 7; an alternative legitimate definition could produce different cluster indices and invalidate the uniqueness claim. Additionally, the 20 repeated runs per case (Section 3.2) are almost certainly deterministic IDM simulations with no stochastic perturbation, so the 'without exception' consistency may only reflect a single trajectory replayed 20 times, not robust evidence of consistent recognition across varied conditions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an online evolving Finite State Machine (e-FSM) for automated vehicle control. The e-FSM uses evolving Takagi-Sugeno (eTS) online clustering to determine states uniquely as new observations arrive, computes a probability distribution over recognized states via a similarity function, and learns action-conditioned transition matrices using the online stochastic method of Filev and Kolmanovsky. The framework is illustrated in a SUMO car-following scenario with IDM controllers of different aggressiveness, run 20 times per case for four controller-switch cases. The authors claim that a collision-prone Dead-End state is uniquely determined and consistently recognized as state #3, and that one-step-ahead state predictions, validated by Jensen-Shannon divergence (JSD) below 0.15, show the transition matrices are precisely identified.","tokens_in":11093,"tokens_out":3144,"duration_ms":33998,"significance":"If the claims hold, the e-FSM would be a useful online component for detecting latent driving risks and predicting future states, with a concrete mechanism for evolving the state set without a fixed reward function or pre-enumerated scenarios. The paper provides explicit equations for clustering, recognition, and transition learning, and it makes a falsifiable prediction about the repeatability of state identification. However, two load-bearing issues currently prevent the claims from being accepted as demonstrated: the recognition probabilities depend on an undefined quantity in the similarity function, and the transition-validation protocol is in-sample. The paper's strengths are its clear formulation of the online learning problem and its explicit experimental setup, but the evidence as presented is not yet reproducible.","major_comments":[{"comment":"The similarity function η_i^t(z_t) = exp( -(z_t - z*_i)^T (z_t - z*_i) / var(z*_i) ) uses the quantity var(z*_i), the variance of a cluster center, which is never defined anywhere in the manuscript. A cluster center is a single vector, so its variance is not well-defined without an explicit convention such as the variance of the observations assigned to that cluster, a per-dimension variance, or a user-specified scalar. Because Prob(S_t) is computed from this equation, and because the transition-matrix updates (Eqs. 15–17), the predicted distributions (Eq. 18), and the JSD results in Figure 5 all depend on Prob(S_t), the experimental evidence cannot be independently reproduced. The claim that the Dead-End state is uniquely and consistently recognized as state #3 may be an artifact of an unreported implementation choice for this variance. The authors should define var(z*_i) precisely, state how it is computed, and report the resulting state-recognition results.","section":"§2.2.2, Eq. (7)"},{"comment":"The validation of the transition matrices is entirely in-sample. The transition matrix Pa(r)_t is updated online using τ(t) = Prob(S_{t-1}) and γ(t) = Prob(S_t) from the same simulation run, and then Prob_pred(S_{t+1}) = Pa(r)_t · Prob(S_t) is compared with the same run's Prob(S_{t+1}). Low JSD therefore measures how well the model fits the exact data used for its own updates, not how precisely the state-transitions are identified in any predictive sense. To support the claim that the transition matrices are precisely identified, the authors should perform out-of-sample validation, for example by training on the first portion of each trajectory and testing on held-out later portions, by cross-validating across the 20 runs, or by comparing the JSD against a baseline model such as the empirical marginal transition frequencies.","section":"§3.2.2, Figure 5"},{"comment":"The paper states that each of the four cases is simulated 20 times, giving 80 simulations in total, but no source of stochasticity is specified anywhere in the experimental setting. The SUMO and IDM models are deterministic given the initial conditions and parameters, so it is likely that the 20 repetitions are identical replays of the same trajectory. If so, the statement that state #3 is recognized 'without exception' in all 80 simulations only demonstrates self-consistency of a single deterministic run, not consistency across varied conditions. The authors should either introduce explicit stochastic perturbations (e.g., in initial speeds, IDM parameters, or observation noise) and report the distribution of recognized states across runs, or explicitly state that the repetitions are identical and temper the consistency claim accordingly.","section":"§3.2, experiment process"}],"minor_comments":[{"comment":"The initialization of the potential recursion is incomplete: the text says b1 is initialized to 0 and the first input is set as a cluster center with potential 1, but Equation (4) defines b_t = b_{t-1} + z_{t-1}^T z_{t-1}, which needs a specification for t=1 and a consistent treatment of z_0.","section":"§2.2.1, Eq. (4)"},{"comment":"The notation for cluster centers is inconsistent across equations: Equation (6) uses z*_t^i, while Equation (7) uses z*_i and later text uses z*_i^t. The authors should standardize the superscript/subscript convention to avoid ambiguity.","section":"§2.2.2, Eq. (7)"},{"comment":"The learning rate φ in Equations (13)–(17) is never given a numerical value. Since the transition-matrix identification results depend on this rate, the authors should report the value used in the simulations.","section":"§3.1, experimental settings"},{"comment":"The eTS parameters ε=0.3 and ρ=0.85 are described as arbitrarily assigned, which is acceptable for a demonstrative study, but the paper should include a sensitivity analysis or at least a discussion of how the number of states and the recognition of state #3 depend on these thresholds.","section":"§3.1, experimental settings"},{"comment":"The action discretization description appears inconsistent: the continuous action set Ac = [−2.5, 2.5] m/s^2 is said to be encoded with a range of 0.3 m/s^2 into 17 intervals, but the arithmetic (5.0/0.3 ≈ 16.67) does not produce exactly 17 intervals. The exact bin edges and whether the endpoints are inclusive should be clarified.","section":"§3.1, action encoding"},{"comment":"The claim that JSD values below 0.15 indicate accurate prediction would be strengthened by a stated threshold or a comparison with a null distribution; without such context, 'significant similarity' is asserted rather than demonstrated.","section":"§3.2.2, Figure 5"},{"comment":"The action reviser module is introduced but its operation is not described; the paper focuses on e-FSM, so this is acceptable as a future-work direction, but the authors should mark it explicitly as outside the scope of the current validation.","section":"§4, framework overview"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a well-structured proposal with a concrete algorithm and a clear experimental setup, but the undefined variance in Eq. (7) is a genuine reproducibility barrier that the authors must fix before the results can be trusted. The in-sample validation and the likely deterministic nature of the 20 repetitions further weaken the empirical support for the central claims. The paper would benefit from the authors making the code and parameter settings publicly available, since the current text leaves several implementation details ambiguous. I recommend major revision rather than rejection because the underlying idea is plausible and the identified issues are fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this before reading: the paper is a modest extension of the authors' own prior work (Filev-Kolmanovsky generalized Markov models plus eTS clustering), with one genuinely new twist: action-conditioned transition matrices. The application to latent-risk state recognition in automated vehicles is fresh, but the machinery underneath is established. If you are expecting a fundamentally new method, you will be disappointed; if you are open to an incremental but plausible combination, this is a reasonable read.\n\nWhat the paper does well: it lays out the e-FSM framework clearly, explains how states are determined online via eTS and recognized via probability distributions, and demonstrates, in a simple car-following setup, that a collision-prone Dead-End state is consistently mapped to a single cluster (state #3). The JSD values below 0.15 between predicted and recognized state distributions are a positive signal, though not a strong one. The writing is straightforward and the experimental setup is easy to follow.\n\nNow the soft spots, in proportion. The stress-test note is correct: Eq. 7 uses var(z*_i) without defining it. A cluster center is a vector, so its variance is not well-defined without an explicit convention (e.g., variance of the observations assigned to that cluster). Since every probability distribution in the paper flows through this equation, the experimental results are not reproducible as written. That is a real blocker, not a nit. Second, the 20 repeated runs per case appear to be deterministic IDM simulations with no stochastic perturbation, so \"without exception\" may just mean the same trajectory was replayed 20 times. That weakens the consistency claim considerably. Third, the transition-prediction evaluation is in-sample: the same recognized state sequence used to update the matrices is used to test the one-step-ahead predictions, so low JSD can reflect fitting rather than predictive accuracy. Fourth, no sensitivity analysis is reported for the arbitrary parameters (epsilon, rho, phi, delta), and the safety framework's \"action reviser\" is only sketched, not implemented. The claim about \"choosing a better action\" is therefore not actually demonstrated.\n\nOverall, the central idea is sound: online evolving state recognition coupled with action-conditioned Markov transitions is a plausible building block for AV decision-making. The paper is not a breakthrough, but it is a coherent, honest piece of work with one serious omission (the undefined variance) and several addressable weaknesses. A serious editor should send it to peer review, expecting major revision. I would not cite it in my own work yet, but I would bring it to a reading group to discuss what needs to be fixed before the claims can be trusted.","headline":"Modest extension of the authors' own prior work, with a real reproducibility gap (undefined variance in Eq. 7) that undercuts the headline claims, but the core idea is worth refereeing.","tokens_in":11593,"tokens_out":1998,"would_cite":false,"duration_ms":23769,"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":"An online-evolving finite state machine can learn a collision-prone Dead-End state from scratch and predict one-step-ahead driving states accurately in simulations.","keywords":["evolving finite state machine","online clustering","latent risk detection","automated vehicle safety","Markov chain","car-following simulation","Jensen-Shannon divergence","Dead-End state"],"falsifier":"Repeat the same car-following experiment with different eTS thresholds (for example epsilon=0.1 or epsilon=0.5) or with a differently scaled observation vector; if collision events then map to more than one state index, or the Jensen-Shannon divergence between predicted and recognized states exceeds 0.15 repeatedly, the paper's uniqueness and precision claims would be contradicted.","tokens_in":10551,"feed_emoji":"🚗","tokens_out":5682,"duration_ms":57754,"temperature":0.7,"pith_summary":"This paper proposes the evolving Finite State Machine (e-FSM), a model that builds its own set of driving states online from sensor observations and learns how those states transition under each available action. The authors aim to show that this lets an automated vehicle controller recognize situations it was never pre-programmed to expect, including a collision-prone 'Dead-End' state, and choose safer actions. In repeated car-following simulations with different driver styles, they report that every collision is recognized as the same state (state #3), and that one-step-ahead state probabilities predicted from the learned transition matrices match the observed recognition within a Jensen-Shannon divergence below 0.15. If the claim holds, an AV controller could detect latent risks online and compare candidate actions by their predicted future states, without a hand-designed list of dangerous situations.","feed_headline":"Evolving state machine spots latent collision states online","feed_subtitle":"The same Dead-End state appears in every collision; predicted states match recognized ones within 0.15 JSD.","key_machinery":"The load-bearing object is the evolving Finite State Machine (e-FSM): a growing set of states, each represented by the center of a cluster of observations, plus one transition matrix for each discrete action. A new state is created when the evolving Takagi-Sugeno online clustering rule decides the current observation has higher potential than all existing cluster centers and is farther than threshold epsilon from the nearest center; otherwise the observation is assigned to clusters by a similarity function that produces a probability distribution over states. Each transition matrix is updated by a stochastic approximation rule using the previous and current state probability distributions, and its dimension grows whenever a state is added. The mechanism that carries the argument is the combination of these two devices: clustering turns continuous observations into discrete states online, and action-conditioned Markov matrices make one-step-ahead prediction a simple matrix-vector product.","core_discovery":"The central claim is that an online-evolving state machine can determine, recognize, and predict driving-risk states without a predefined state set. States are cluster centers of observations (headway, follower speed, preceding-vehicle speed) created by an online clustering rule; recognition is a probability distribution over these clusters; and state transitions are stored in one Markov transition matrix per discrete action, updated online and expanded whenever a new state appears. Using an Intelligent Driver Model controller in a simulated car-following scenario with four controller configurations repeated 20 times, the paper reports that a Dead-End state—the situation in which a collision is no longer avoidable—is always recognized as state #3, and the Jensen-Shannon divergence between predicted and recognized one-step-ahead state distributions is less than 0.15 except for the very first prediction. The paper concludes that the Dead-End state is uniquely determined and consistently recognized and that the identified transition matrices are accurate enough for future-state prediction.","pith_inferences":["The paper does not specify how the cluster variance in the similarity function is computed; a natural extension would be to test whether the uniqueness of the Dead-End state survives alternative variance definitions or normalized observations.","The stability of the result under different eTS thresholds is untested; changing epsilon or rho could produce a different partition and possibly split or merge the collision state, so a threshold sensitivity study would delimit the claim.","The observation vector contains only longitudinal quantities; extending e-FSM to lateral position, lane context, or map information would test whether the same unique-state property holds in more realistic traffic situations.","If the approach scales, the same mechanism could label risky states in other safety-critical systems, such as robotic manipulation or human-robot interaction, where unanticipated states must be recognized online rather than enumerated."],"forward_implications":["An AV controller can be given an early warning whenever the current observation matches the learned Dead-End state, before the collision actually occurs.","Because each action has its own transition matrix, the controller can compare candidate actions by the future-state distributions they imply, rather than only by immediate reward.","The state set is not fixed in advance: the model can keep adding states as new situations appear, and in the reported experiments the number of states stabilizes at 7 after four simulations.","The same recognized state index can serve as a shared risk label across different driver styles and controller settings, since the identical unsafe situation maps to state #3 in all cases."],"supporting_citations":[{"why":"Supplies the online Markov-chain transition identification equations that e-FSM adapts into action-conditional transition matrices.","marker":"(13)"},{"why":"Provides the evolving Takagi-Sugeno clustering rule that decides when an observation becomes a new state (cluster center).","marker":"(14,15)"},{"why":"Supplies the Intelligent Driver Model used as the vehicle controller in the simulated car-following experiments.","marker":"(16,17)"},{"why":"Provides the Jensen-Shannon divergence measure used to compare predicted and recognized state distributions.","marker":"(18)"}],"fun_headline_variants":["Evolving state machine uniquely spots collision danger state","Online evolving FSM recognizes latent collision states","State machine predicts collision risk via online evolving model","Evolving FSM identifies collision states with 0.15 JSD accuracy","Dead-End state uniquely recognized by online evolving FSM"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the online clustering rule, with its arbitrarily chosen thresholds epsilon=0.3 and rho=0.85 and an unspecified cluster variance, partitions observations so that each cluster corresponds to one unique driving situation; the claim that every Dead-End event maps to state #3 depends entirely on that partition.","fun_headline_variants_meta":{"raw":{"variants":["Evolving state machine uniquely spots collision danger state","Online evolving FSM recognizes latent collision states","State machine predicts collision risk via online evolving model","Evolving FSM identifies collision states with 0.15 JSD accuracy","Dead-End state uniquely recognized by online evolving FSM"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000516,"raw_usage":{"total_tokens":2528,"prompt_tokens":994,"completion_tokens":1534,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":1457}},"tokens_in":610,"tokens_out":1534,"duration_ms":12789,"temperature":1.0,"reasoning_tokens":1457,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:33:24.637049+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the same car-following experiment with different eTS thresholds (for example epsilon=0.1 or epsilon=0.5) or with a differently scaled observation vector; if collision events then map to more than one state index, or the Jensen-Shannon divergence between predicted and recognized states exceeds 0.15 repeatedly, the paper's uniqueness and precision claims would be contradicted.","supporting_citations":[],"review_version":1}