{"id":"5d485465-8e3e-4629-92c8-352032a2b5d0","arxiv_id":"2505.06017","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Adaptive-UCS evolves a per-rule binary fuzzy indicator that switches between rectangular and triangular membership functions, improving average classification accuracy over crisp UCS and fuzzy-hypertrapezoidal Fuzzy-UCS on most tested problems.","lead":"The paper adds a learnable flag to each rule in a fuzzy learning classifier system, letting the system choose between crisp rectangular and fuzzy triangular membership per input dimension. Tests on benchmarks and 20 real-world datasets report higher average accuracy and robustness to noise and missing values than two prior UCS variants.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Self-adaptation claim is confounded: Adaptive-UCS changes theta_overlap (0.5 vs 0.8) and the representation simultaneously, and no ablation fixes the fuzzy indicator, so the accuracy gains are not yet attributable to evolved rule shapes.","rationale":"The reader's weakest assumption identifies the theta_overlap change as a comparison-fairness risk, which is valid and load-bearing. I agree with that concern and would add a second, closely related one: even if theta_overlap were held fixed, the paper lacks an ablation that fixes the fuzzy indicator F, so the specific contribution of self-adaptation is not isolated from the simpler two-parameter representation or from the mixture of crisp and fuzzy rules. The reader's CONDITIONAL verdict already captures the appropriate level of caution: the proposal is plausible and the experiments are broad, but the central attribution claim is not yet established. Therefore I do not move the verdict. My partial agreement reflects that the reader focused on hyperparameter fairness while I see the missing F-fixed ablation as at least as central to the self-adaptation claim.","tokens_in":21516,"tokens_out":7236,"duration_ms":78638,"concrete_test":"On RCB, NCB, and the 20 UCI datasets, run a 2x2 sweep of theta_overlap in {0.5, 0.8} for Adaptive-UCS and Fuzzy-UCS with all other hyperparameters fixed, and additionally run Adaptive-UCS variants with F fixed to all 0, all 1, and random-but-not-evolved. If Fuzzy-UCS at theta_overlap = 0.5 matches Adaptive-UCS's advantage, or if a fixed-F variant matches the evolved-F variant, then the reported superiority is not evidence for self-adaptation of rule representations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that evolving the fuzzy indicator F is what improves accuracy, but the experiments never isolate that mechanism. Adaptive-UCS differs from Fuzzy-UCS in two ways at once: (i) it uses a simpler center-spread two-parameter representation whose fuzzy mode is a degenerate triangle, and (ii) Section 5.2 sets theta_overlap = 0.5 instead of Fuzzy-UCS's 0.8. Because theta_overlap directly controls the subsumption/generality trade-off in Eq. (8), lowering it permits more aggressive deletion of less general rules, and Table 1 indeed shows Adaptive-UCS consistently producing smaller rulesets than Fuzzy-UCS. Without a theta_overlap sweep on all systems, or an ablation that fixes F to all-crisp, all-fuzzy, or randomly initialized but non-evolved configurations while keeping the rest of Adaptive-UCS unchanged, the reported accuracy gains cannot be attributed to self-adaptation of rule shapes rather than to the changed subsumption threshold or to the simpler representation. The paper's own admission that F mutation is only random (Section 6.2) and that Adaptive-UCS underperforms on some test datasets further weakens the abstract's unqualified 'outperforms' claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Adaptive-UCS is a Michigan-style supervised learning classifier system that extends Fuzzy-UCS by adding a binary fuzzy indicator F_i to each rule condition dimension. Each dimension is represented by a center c_i and spread s_i, and F_i selects either a rectangular (crisp) or triangular (fuzzy) membership function. The indicator is inherited, crossed over, and mutated along with the continuous parameters. The paper compares Adaptive-UCS with UCS (crisp hyperrectangles) and Fuzzy-UCS (hypertrapezoids) on three synthetic checkerboard problems (CB, RCB, NCB) and 20 real-world datasets, reporting accuracy and ruleset size from 30 runs with ANOVA/Welch and post-hoc tests. The central claim is that Adaptive-UCS outperforms both baselines in classification accuracy and is robust to noise and missing values.","tokens_in":21773,"tokens_out":4452,"duration_ms":44790,"significance":"Validating the core claim would be useful: automatic per-rule selection between crisp and fuzzy conditions addresses a known limitation of fixed rule representations in LCSs, and the two-parameter condition keeps the search space smaller than the hypertrapezoidal representation. The paper's empirical discipline is a strength: 30 independent trials, multiple benchmarks, and appropriate parametric/non-parametric statistical procedures. However, the experiments do not currently isolate the proposed mechanism from a changed comparison hyperparameter, and the abstract's unqualified superiority statement is stronger than the per-dataset results in Table 2. The contribution is potentially publishable, but the evidence needs strengthening and the claims need qualification.","major_comments":[{"comment":"Section 5.2 sets Adaptive-UCS identical to Fuzzy-UCS \"except theta_overlap = 0.5\", while Fuzzy-UCS uses 0.8. Because Eq. (8) uses theta_overlap to decide whether a candidate subsumer is more general, the two systems differ in subsumption and generality pressure, not only in the adaptive membership function. Table 1 shows that Adaptive-UCS yields smaller rulesets than Fuzzy-UCS in all three benchmark problems, which is consistent with a changed subsumption threshold. To attribute the accuracy gains to evolving F, the authors should either run all systems under a sweep of theta_overlap, or add ablations of Adaptive-UCS with F fixed to all-crisp, all-fuzzy, and random-but-static values under otherwise identical settings.","section":"5.2, Eq. (8)"},{"comment":"The abstract's claim that Adaptive-UCS \"outperforms other UCSs ... in classification accuracy\" is not supported as a blanket statement. In Table 2, Adaptive-UCS test accuracy is lower than UCS on Horse colic (59.63 vs 61.39) and lower than Fuzzy-UCS on Soybean (59.90 vs 68.48) and Wisconsin prognostic breast-cancer (68.25 vs 72.81), and Section 6.2 explicitly acknowledges underperformance on some datasets. The claim should be restricted to average rank or to the majority of datasets, with the exceptions reported in the abstract and conclusion.","section":"Abstract; Table 2"},{"comment":"The mechanism is advertised as self-adaptive, but Eq. (7) flips F_i with no dependence on any local gradient or feedback, and the paper never reports the distribution of evolved F values or compares against non-evolved F baselines. As a result, the paper does not currently demonstrate that evolutionary optimization of F, rather than the changed representation palette alone, is responsible for the reported gains. An analysis of the final F distribution per problem, together with the ablation described in the first major comment, would close this gap.","section":"4.4.2, Eq. (7); 6.2"}],"minor_comments":[{"comment":"The text \"One-Way ANOV Aand Tukey-HSD\" contains a typo: \"ANOVA\" and the spacing should be corrected; also, \"homoscedasticity was positive/negative\" would be clearer as \"the null hypothesis of equal variances was rejected/not rejected\".","section":"5.2"},{"comment":"The phrase \"frequently need help addressing problems\" is awkward and likely means \"often struggle to address\"; consider rewording.","section":"Abstract"},{"comment":"The color shading used to indicate best and worst groups may not survive grayscale printing; adding textual markers such as bold plus asterisks would improve accessibility.","section":"Tables 1 and 2"},{"comment":"The condition \"no cross. has taken place\" is abbreviated; spell out \"no crossover has taken place\" for clarity.","section":"4.4.2, Eq. (6)"},{"comment":"The paper does not state whether the source code is available; sharing code would aid reproducibility and is encouraged.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has clear local strengths, but the missing theta_overlap sweep and F-ablations are essential for the central attribution claim. If the authors add those experiments and qualify the abstract appropriately, I would support acceptance. I see no concerns about novelty or authorship."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:2505.06017.\n\nAdaptive-UCS is a genuine, modest extension of Fuzzy-UCS. It adds a binary fuzzy indicator F_i to each rule condition, parameterizes each condition as center/spread, treats F=0 as a rectangle and F=1 as a triangle, and evolves F through crossover and mutation. The matching, covering, and subsumption operators are all adapted consistently. That is a concrete new mechanism within the UCS line, and the paper explains it clearly. The experimental work is careful in procedure: 30 trials, appropriate parametric/nonparametric tests, clear tables, and boundary visualizations that help the reader see what the learned rules look like.\n\nThe soft spot is attribution. The stress-test note is right. Adaptive-UCS differs from Fuzzy-UCS in two ways at once: the center-spread representation and the choice theta_overlap=0.5 vs 0.8. Since theta_overlap appears directly in the is-more-general condition of Eq. (8), it changes how aggressively rules subsume. No ablation isolates the fuzzy indicator—fixing it to all-crisp, all-fuzzy, or random-but-not-evolved would have been straightforward and decisive. Without that, the accuracy gains cannot be confidently credited to self-adaptation of rule shapes. This is not a manufactured flaw; the paper's own Table 1 shows Adaptive-UCS producing smaller rule sets, which the lower theta could explain.\n\nSecond, the abstract overclaims. It says Adaptive-UCS outperforms the other UCSs, but Table 2 shows several test datasets where it is numerically worse: Horse colic, Pistachio, Raisin, Soybean, and Wisconsin prognostic breast-cancer. Average rank is in its favor and many differences are statistically significant, but blanket superiority is not the result. The body is more honest than the abstract, mentioning underperformance on a few datasets.\n\nThird minor: no code is provided, and the fuzzy indicator mutation is a random bit-flip, so calling the mechanism 'self-adaptive' is generous in the sense that it is simply evolved. The idea is still sound, but it would benefit from a sentence saying why random search over a single bit is the intended mechanism.\n\nWho should read it: people working on LCS representations or interpretable fuzzy rule systems. It is a reasonable follow-up to Fuzzy-UCS and worth citing. It deserves peer review, not a desk reject, but I would send it back for major revision: add the ablation, run a theta_overlap sweep across all systems, and rewrite the abstract to match the table. If the results survive that, it becomes a solid contribution. The duplicated text in the extraction looked like a PDF artifact, not a scientific issue.","headline":"A genuine but modest extension of Fuzzy-UCS whose headline claim runs ahead of the experiments: the gains are real on average, but the self-adaptation mechanism is not isolated from a changed theta_overlap and a simpler representation.","tokens_in":22270,"tokens_out":3611,"would_cite":true,"duration_ms":37877,"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":"This paper proposes Adaptive-UCS, a Fuzzy-UCS variant in which each rule condition carries an evolved fuzzy indicator that switches between crisp rectangular and fuzzy triangular membership, and claims this self-adapting representation…","keywords":["Learning Fuzzy-Classifier Systems","Fuzzy-UCS","Adaptive-UCS","self-adaptive rule representation","fuzzy indicator","crisp-hyperrectangular rules","fuzzy-hypertrapezoidal rules","supervised learning"],"falsifier":"Train Adaptive-UCS with every fuzzy indicator frozen to 0 (crisp-only) and, separately, frozen to 1 (fuzzy-only) on the same checkerboard and real-world datasets. If either frozen version matches the full Adaptive-UCS accuracy, the reported gains come from the fixed representation rather than from evolving the indicator. Alternatively, rerun Fuzzy-UCS with $\\theta_{\\text{overlap}}$ set to $0.5$ instead of $0.8$; if it closes the gap, the relaxed subsumption threshold, not shape self-adaptation, explains the difference.","tokens_in":21294,"feed_emoji":"🧬","tokens_out":10838,"duration_ms":100725,"temperature":0.7,"pith_summary":"The paper sets out to establish that a Michigan-style learning classifier system does not have to commit in advance to either crisp or fuzzy rules. Adaptive-UCS adds a Boolean fuzzy indicator to each rule condition; when the indicator is 0 the condition is a rectangular interval, when it is 1 it is a triangular fuzzy set, and the indicator is inherited, crossed over, and mutated along with the rule's numeric parameters. On three synthetic checkerboard-style benchmarks and twenty real-world classification datasets, the paper reports that Adaptive-UCS attains higher average test accuracy than both conventional crisp-hyperrectangular UCS and fuzzy-hypertrapezoidal Fuzzy-UCS, with no penalty in ruleset size. It also reports stable accuracy when inputs are noisy or attributes are missing. If the claim holds, users of such systems would no longer have to guess whether a problem wants crisp or fuzzy rule shapes.","feed_headline":"Rules evolve their own crisp or fuzzy shape to classify better","feed_subtitle":"Each rule learns whether its conditions should be crisp or fuzzy, beating both fixed rule shapes.","key_machinery":"The load-bearing mechanism is the fuzzy indicator $\\mathcal{F}_i$ attached to every condition of every rule, combined with the center-spread encoding $A_i=(c_i,s_i)$. The indicator selects the membership function shape, and because it is subject to the same crossover and mutation as the numeric parameters, the evolutionary process explores which dimensions should be crisp and which should be fuzzy. Around this sits the rest of the Fuzzy-UCS machinery, including match sets, correct sets, fitness based on class weights, the accuracy-based genetic algorithm, and a subsumption-with-merge operator whose is-more-general test uses an overlap ratio $\\theta_{\\text{overlap}}$ when either condition is fuzzy. The result is that a single rule template can represent up to $2^n$ combinations of crisp and fuzzy regions.","core_discovery":"The central discovery, on the paper's own terms, is that the representational choice between crisp and fuzzy rule conditions can be optimized by the same evolutionary loop that learns the rules. Adaptive-UCS represents each condition by a center and spread, $A_i=(c_i,s_i)$, plus a fuzzy indicator $\\mathcal{F}_i\\in\\{0,1\\}$: with $\\mathcal{F}_i=0$ the membership function is rectangular, giving an all-or-nothing match inside $[c_i-s_i,c_i+s_i]$; with $\\mathcal{F}_i=1$ it is an isosceles triangle peaking at $c_i$, giving a graded match. The fuzzy indicator is crossed over and mutated along with $c_i$ and $s_i$, so a single ruleset can contain crisp and fuzzy conditions in any mixture and the system itself finds the mixture a problem needs. The paper's experiments show Adaptive-UCS at the best average rank on both synthetic and real-world benchmarks, with its largest margins on problems where class boundaries are oblique, inputs are noisy, or attributes are missing. The authors conclude that self-adapting two simple membership shapes can beat both fixed representations.","pith_inferences":["Beyond the paper: freezing every fuzzy indicator to 0 or to 1 would isolate how much of the gain comes from the evolutionary search over shapes; the paper reports only the full self-adaptive system, so this ablation is a natural test.","Beyond the paper: the per-dimension indicators could be read after training as a map of where a dataset actually needs graded boundaries, turning the mechanism into a diagnostic tool for feature geometry.","Beyond the paper: the reported accuracy gap may partly reflect the relaxed $\\theta_{\\text{overlap}}=0.5$ subsumption threshold used only for Adaptive-UCS; a controlled sweep over $\\theta_{\\text{overlap}}$ for both systems would separate the representation effect from the subsumption effect.","Beyond the paper: extending the indicator alphabet to more than two shapes, such as trapezoids or Gaussian membership functions, is not covered by the experiments, and the paper's argument that simpler shapes suffice is a conjecture about where the complexity-optimizability trade-off sits."],"forward_implications":["If the central claim is correct, a designer no longer needs to choose between crisp and fuzzy rule representations before training; Adaptive-UCS can search that choice itself for a given dataset.","On problems with mixed geometry, where some features have sharp boundaries and others have gradual ones, self-adaptation should beat any single fixed shape, which is exactly the pattern the paper reports on oblique-boundary and noisy benchmarks.","The ruleset-size results in the paper indicate that representational flexibility does not come at the price of larger or more redundant rule sets.","Because the fuzzy indicator is independent of the learning task, the mechanism is claimed to transfer to other learning classifier systems, such as XCS for reinforcement learning or XCSF for function approximation.","For data with noise or missing values, the paper implies that the system can fall back on fuzzy conditions where needed and crisp conditions elsewhere, giving stable test accuracy instead of overfitting."],"supporting_citations":[{"why":"Base Fuzzy-UCS algorithm that Adaptive-UCS extends; supplies the training/test loop, fitness update, and genetic algorithm.","marker":"[32]"},{"why":"Fuzzy-hypertrapezoidal rule representation and merge mechanism used as the comparison baseline and adapted into Adaptive-UCS.","marker":"[38]"},{"why":"Crisp-hyperrectangular rule representation used by the UCS baseline for continuous-valued inputs.","marker":"[42]"},{"why":"Introduces the fuzzy hyper-trapezoidal membership functions behind the is-more-general overlap test.","marker":"[37]"},{"why":"Center-spread hyperrectangular representation that Adaptive-UCS's crisp mode is equivalent to.","marker":"[52]"},{"why":"Provides most of the twenty real-world datasets used in the second experiment.","marker":"[15]"},{"why":"Source of the standard hyperparameter configuration applied to all three systems.","marker":"[45]"},{"why":"Supplies the noisy-environment parameter settings used in the noisy checkerboard experiment.","marker":"[47]"}],"fun_headline_variants":["Self-adapted rule shapes beat fixed crisp or fuzzy","Evolution finds optimal rule shape: crisp or fuzzy","When to be crisp, when to be fuzzy: rules evolve","Each rule self-selects crisp or fuzzy membership","Rules evolve when to be crisp or fuzzy on their own"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"That the comparison is fair: all three systems use hyperparameters taken from earlier papers, and Adaptive-UCS's one described deviation, $\\theta_{\\text{overlap}}=0.5$ instead of Fuzzy-UCS's $0.8$, is not itself responsible for the accuracy gains.","fun_headline_variants_meta":{"raw":{"variants":["Self-adapted rule shapes beat fixed crisp or fuzzy","Evolution finds optimal rule shape: crisp or fuzzy","When to be crisp, when to be fuzzy: rules evolve","Each rule self-selects crisp or fuzzy membership","Rules evolve when to be crisp or fuzzy on their own"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000956,"raw_usage":{"total_tokens":4092,"prompt_tokens":981,"completion_tokens":3111,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":3034}},"tokens_in":597,"tokens_out":3111,"duration_ms":22268,"temperature":1.0,"reasoning_tokens":3034,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:50:13.724068+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Adaptive-UCS with every fuzzy indicator frozen to 0 (crisp-only) and, separately, frozen to 1 (fuzzy-only) on the same checkerboard and real-world datasets. If either frozen version matches the full Adaptive-UCS accuracy, the reported gains come from the fixed representation rather than from evolving the indicator. Alternatively, rerun Fuzzy-UCS with $\\theta_{\\text{overlap}}$ set to $0.5$ instead of $0.8$; if it closes the gap, the relaxed subsumption threshold, not shape self-adaptation, explains the difference.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Base Fuzzy-UCS algorithm that Adaptive-UCS extends; supplies the training/test loop, fitness update, and genetic algorithm."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Fuzzy-hypertrapezoidal rule representation and merge mechanism used as the comparison baseline and adapted into Adaptive-UCS."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Crisp-hyperrectangular rule representation used by the UCS baseline for continuous-valued inputs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the fuzzy hyper-trapezoidal membership functions behind the is-more-general overlap test."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Center-spread hyperrectangular representation that Adaptive-UCS's crisp mode is equivalent to."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the standard hyperparameter configuration applied to all three systems."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the noisy-environment parameter settings used in the noisy checkerboard experiment."}],"review_version":1}