{"id":"acfce3e3-97f7-4591-8691-14db73ea5feb","arxiv_id":"2412.03710","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"CIKAN, a KAN-based constraint-informed network, approximates the Time Shift Governor for spacecraft rendezvous and, in simulation, enforces constraints while reducing average computation time and fuel use relative to the conventional TSG.","lead":"This paper trains Kolmogorov-Arnold neural networks to predict the time-shift parameter used by a Time Shift Governor, an add-on controller that keeps spacecraft rendezvous trajectories inside safety constraints. Simulations show these networks run faster on average than solving the optimization online and use far fewer parameters than multilayer perceptron alternatives.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Speed claim is uninterpretable without fallback-frequency statistics for Algorithm 1; the candidate acceptance rate is the missing quantity.","rationale":"The central claim is that CIKAN-based TSG outperforms conventional TSG in average computation time and fuel consumption while enforcing constraints. For the computation-time part, the hybrid Algorithm 1 means the measured average depends on how often the exact solver is used as a fallback. The reader's weakest assumption identifies exactly this: the claimed speed advantage depends on Algorithm 1 being dominated by NN evaluations, and the paper reports no fallback statistics. I agree that this is the most load-bearing concern. The line-5 feasibility check itself also has a cost, but it is not separately reported; even a single forward simulation per step could dominate the NN evaluation cost, so the speedup should be decomposed into NN inference, feasibility-check simulation, and any exact-solver fallback. The dataset representativeness concern is real but secondary: because of the fallback, poor coverage mainly increases fallback frequency rather than causing silent constraint violations, as long as the feasibility check is correct. I additionally note a likely sign error in Eq. (8), where exp(·) gives a positive shift despite the stated non-positive focus; this should be fixed in a revision but does not change the main verdict because the reported simulation results appear to reflect the intended negative shifts. The appropriate disposition is therefore the same CONDITIONAL verdict the reader gave: the paper is plausible and addresses an interesting application, but its headline quantitative claims need the missing statistics and artifact release before they can be trusted.","tokens_in":9561,"tokens_out":5004,"duration_ms":49121,"concrete_test":"Instrument Algorithm 1 in the released or reimplemented code to record, for each Monte Carlo and extreme-case run, the fraction of discrete-time steps where lines 9 or 11 invoke π*, the wall-clock time spent in the candidate feasibility check, and the time spent in π*. Then recompute Table 2's average computation time using only the steps where the candidate is accepted without fallback. If the candidate acceptance rate is below, say, 90%, or if the candidate-only average time is not below the TSG average, the headline speedup is not attributable to the NN.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 (lines 4-11) accepts the CIKAN prediction only if it lies in the previous feasible interval and passes a forward-simulation feasibility check; otherwise it calls the exact TSG solver π*. The paper reports average computation time 0.0329 s for CIKAN in Table 2 but never reports the fraction of steps in which the fallback to π* is invoked. If that fraction is high, the measured average is dominated by the exact solver, and the claimed speed advantage over conventional TSG (0.0871 s) becomes an artifact of the hybrid design rather than of the KAN approximation. The same issue affects the Delta V comparison: trajectories produced by different mixtures of candidate and fallback calls are different maneuvers, so fuel savings cannot be attributed to CIKAN alone. The dataset representativeness issue is secondary because the line-5 feasibility check is a safety net; however, out-of-distribution states will exercise the fallback, which is precisely the quantity that needs reporting. Without per-step acceptance rates and per-seed variance, the central quantitative claim is unverifiable. There is also a presentation inconsistency: Eq. (8) as written returns a positive time shift from exp(·), while the paper states the focus is on non-positive shifts and Algorithm 1 constrains T to non-positive values; the figures show negative shifts, suggesting a missing minus sign, and this discrepancy should be resolved before the algorithm can be assessed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CIKAN, a Kolmogorov-Arnold Network based constraint-informed approximation of the Time Shift Governor (TSG) solution mapping for spacecraft rendezvous in a highly elliptic orbit. The authors formulate a log-transformed loss with a constraint penalty, describe a hybrid Algorithm 1 that accepts a neural-network time-shift candidate only when it is feasible and otherwise falls back to the exact TSG solver, train on one million state-to-time-shift pairs generated by the exact TSG, and compare three KAN variants and two MLP variants against the conventional TSG in Monte Carlo and extreme-case simulations. The central claim is that the CINN-based models, especially CIKAN, outperform the conventional TSG in average computation time and fuel consumption while enforcing the line-of-sight, thrust, and approach-velocity constraints.","tokens_in":9793,"tokens_out":4611,"duration_ms":47532,"significance":"If the quantitative claims hold, this is a useful step toward fast approximate implementation of TSG for constrained spacecraft rendezvous, and it is one of the first applications of KANs to a constrained-control add-on scheme. The paper has clear strengths: it benchmarks six models, reports model complexity and validation losses, includes Monte Carlo and extreme-case simulations, and explicitly demonstrates constraint enforcement in Figure 4. However, the main performance claims are not yet statistically supported, and one sign inconsistency in the algorithm description must be resolved.","major_comments":[{"comment":"The central speed claim is not verifiable because the paper never reports how often the CIKAN prediction is accepted in Algorithm 1. Table 2 lists only the average time 0.0329 s and worst-case 0.5590 s for CIKAN; if the fallback call to pi* at line 9 or line 11 fires frequently, the measured average is dominated by the exact TSG solver rather than by the KAN evaluation. Please report the per-step acceptance rate, the computation time separated into accepted-candidate and fallback cases, and the distribution or variance of computation times. Note also that the GRBF-CIKAN worst-case time (1.0339 s) exceeds the conventional TSG worst case (0.8546 s), so the average-time advantage is not a uniform improvement and needs disaggregation.","section":"Algorithm 1; Table 2"},{"comment":"The fuel-consumption comparison is confounded and lacks statistical support. Table 2 reports single Delta V values with no error bars or multiple-seed statistics, despite the Monte Carlo runs over 100 initial states. Because an approximate time shift produces a different closed-loop trajectory than the exact TSG, a lower Delta V can simply reflect a different maneuver rather than a better approximation; the text itself says the approximation errors lead to different state trajectories that lead to improved Delta V performance. To support the conclusion that CIKAN outperforms TSG in fuel consumption, report per-trajectory Delta V statistics, identify which steps used the exact-solver fallback, and separate the effect of the neural approximation from the effect of the hybrid fallback mechanism.","section":"Table 2; Section 4"},{"comment":"There is a sign inconsistency that must be resolved before the method is fully assessable. The paper states it focuses on non-positive time shifts for rendezvous from forward in track, and Algorithm 1 confines T_k to the interval [t_hat_shift,k, 0], yet Eq. (8) defines t_shift = exp(pi_tilde_theta*(x)), which is strictly positive. The figures show negative time shifts, so the implementation apparently uses -exp(...) or the sign convention is opposite to the equations. Please correct Eq. (8) and make the signs in Eqs. (6)-(7), the discussion of non-positive shifts, and Algorithm 1 mutually consistent.","section":"Eq. (7)-(8); Algorithm 1"},{"comment":"The training-data and safety-net descriptions are incomplete for reproducibility and for assessing how much of the constraint satisfaction is due to the neural approximation. The paper does not state how the one million states are sampled, what prediction horizon Np and terminal constraints are used when computing t* via pi*, or how the feasibility check in Algorithm 1 is implemented; line 5 is not a precise if-condition. Since the authors explicitly leave feasibility and convergence guarantees to future work, the simulation evidence is the only support for the method. Without these details and without fallback-frequency statistics, the reader cannot determine whether constraints are enforced by the neural network or by the exact TSG fallback.","section":"Section 4; Eqs. (2)-(4)"}],"minor_comments":[{"comment":"The notation for the constraint penalty weight is inconsistent: Eq. (6) uses theta_CINN, the text says theta_CINN denotes a scalar weight, and Eq. (7) uses theta_c. The same symbol theta is also used for network parameters. Please use distinct symbols for the scalar penalty weight and the parameter vector.","section":"Eq. (6)-(7)"},{"comment":"There is a typo in the sentence 'they complete the spacecraft rendezvous mission is successfully completed'; please rephrase. Also, in Figure 3 the quantity E[t_shift] is not defined; the caption should state explicitly whether the expectation is over the Monte Carlo initial conditions.","section":"Section 4, Figure 3"},{"comment":"The statement that t*_shift(t) is either nonnegative or nonpositive at all times for a specific maneuver is a strong assumption. It should be justified or relaxed, because if the sign of the optimal shift can change, the two separate non-negative and non-positive models would be insufficient.","section":"Section 3.1"},{"comment":"Line 5 reads 'if t_cand_shift ensures constraint satisfaction over a prediction horizon' and is not a complete algorithmic condition. Please specify the exact forward-simulation check, the horizon used, and how the feasibility of the candidate is verified in the implementation.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is in scope for a control-systems venue and the idea is timely. The main risk is that the headline 'outperform' claims rest on missing statistics, particularly the Algorithm 1 fallback acceptance rate. I do not see this as a rejection; the missing quantities are straightforward to report in a revision. I would also encourage the editor to ask the authors for a data/code availability statement, since the training-data generation details are essential for reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper.\n\nFirst, the core idea is legitimate: approximate the Time Shift Governor solution mapping with a KAN, using a log-transformed constraint-informed loss. That combination appears to be new, and the authors run a broad comparison across three KAN variants and two MLP baselines on a realistic elliptic-orbit rendezvous problem.\n\nSecond, the main quantitative claim--CIKAN is faster and uses less fuel than conventional TSG--does not hold up as stated, because Algorithm 1 falls back to the exact TSG solver whenever the candidate fails a feasibility check, and the paper never reports how often that happens. If the fallback fires frequently, the 0.033 s average is not the network's speed, and the Delta-V comparison mixes different maneuvers.\n\nWhat the paper does well: the simulations are coherent, the constraint-satisfaction plots are convincing, and the authors are honest about deferring formal guarantees to future work. The log-transform of the time shift is a sensible trick for learning small values, and the Monte Carlo plus extreme-case setup is a reasonable evaluation. The comparison of model complexity versus validation loss is also informative.\n\nThe soft spots are real but addressable. The missing fallback frequency is the load-bearing one. The paper also reports no error bars or multiple-seed statistics for Table 2, so the differences between CIKAN, CINN-GELU, and the others could be noise. There is a presentation inconsistency in Eq. (8): the paper states a focus on non-positive shifts, but exp(.) produces positive values; a missing minus sign is likely, and it needs to be fixed before the algorithm can be reproduced. No code or data is released, which makes the empirical claims hard to verify.\n\nWho gets value: readers working on neural approximations of predictive or governor-based control laws, and the spacecraft G&C community interested in KAN applications. It is a useful case study, not a breakthrough. I would send it to peer review if the authors can report per-step fallback rates, per-seed variance, and the sign fix. In its current form, I would not cite the quantitative results, but I would cite the method as an example if it survives revision.","headline":"A sound but narrow application of KAN to TSG approximation whose headline speed and fuel numbers are uninterpretable without fallback-frequency statistics.","tokens_in":696,"tokens_out":957,"would_cite":false,"duration_ms":29044,"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":"Replacing the Time Shift Governor's online optimization with a Kolmogorov-Arnold network cuts average computation time and fuel use while still enforcing spacecraft rendezvous constraints.","keywords":["Kolmogorov-Arnold networks","Time Shift Governor","constrained spacecraft rendezvous","line-of-sight constraint","neural network control","elliptic orbit","constraint-informed neural networks"],"falsifier":"Count the number of times Algorithm 1 calls the exact TSG solver over a large battery of Monte Carlo rendezvous runs, and separately disable the fallback and measure line-of-sight cone violations over states outside the training distribution; if the fallback rate is not close to zero, the 0.0329 second average mixes network and optimizer cost, and if violations appear without the fallback, the learned map does not itself enforce the constraints.","tokens_in":9289,"feed_emoji":"🛰️","tokens_out":6145,"duration_ms":56606,"temperature":0.7,"pith_summary":"This paper tries to show that a small constraint-informed neural network, built on the Kolmogorov-Arnold architecture, can replace the repeated online optimization in a Time Shift Governor without losing constraint enforcement. The Time Shift Governor enforces spacecraft rendezvous constraints by shifting the reference trajectory in time; solving this optimization at every step is correct but slow. The authors train a KAN-based network, called CIKAN, on one million state-to-time-shift pairs produced by the exact governor, then run a hybrid algorithm that trusts the network when its prediction is feasible and falls back to the exact solver otherwise. In simulations of rendezvous on a highly elliptic orbit, the CIKAN-based scheme enforces the line-of-sight, thrust, and approach-velocity limits and improves average computation time and fuel use relative to the conventional TSG.","feed_headline":"Neural time-shift governor cuts rendezvous fuel and compute","feed_subtitle":"On a highly elliptic orbit, a Kolmogorov-Arnold network matches the Time Shift Governor's constraints at lower average cost.","key_machinery":"The central object is the scalar time-shift parameter $t_{\\mathrm{shift}}$ of the Time Shift Governor (TSG), an add-on reference governor that enforces constraints by replacing the reference $x_c(t)$ with $x_c(t+t_{\\mathrm{shift}})$ in a nominally stabilizing feedback loop. The optimal shift is the smallest in magnitude shift that keeps the predicted closed-loop trajectory inside the state and input constraint sets, defining the map $\\pi^*$. To carry the learning, the paper uses Kolmogorov-Arnold networks (KANs), whose learnable activation functions are splines on edges, with a loss that combines squared error on $\\log(|t_{\\mathrm{shift}}^*|)$ with a hinge-style penalty that discourages under-prediction of the shift magnitude. Algorithm 1 then wraps the network prediction with a feasibility check and an exact-solver fallback, so the network never directly releases an infeasible shift.","core_discovery":"The central claim is that the mapping $\\pi^*(x(t); x_c(t))$ that returns the optimal time shift can be approximated well enough by a KAN-based constraint-informed network that the approximation, when inserted into the TSG update loop, enforces all three constraints---the line-of-sight cone, the $0.5\\,\\mathrm{m/s^2}$ thrust limit, and the distance-dependent approach-velocity bound---while reducing the average per-step computation from $0.0871\\,\\mathrm{s}$ to $0.0329\\,\\mathrm{s}$ and the average fuel use from $3.8741\\,\\mathrm{km/s}$ to $3.4965\\,\\mathrm{km/s}$ of $\\Delta V$. The authors reformulate the regression target as $\\log(|t_{\\text{shift}}^*|)$ for the non-positive shift case, splitting the problem by sign, and add a constraint-aware penalty to the loss. The trained network is used in Algorithm 1: it predicts a candidate shift, checks feasibility over the prediction horizon, and only calls the exact solver when the candidate is infeasible or outside the previous feasible interval. The paper leaves approximation-error robustness to future work and presents the empirical claim that this hybrid loop succeeds on Monte Carlo and extreme-case scenarios.","pith_inferences":["If the fallback events in Algorithm 1 are rare, the reported average computation time is dominated by a forward pass of a KAN, which suggests that even smaller networks may meet the same accuracy, since GRBF-CIKAN used 122,880 parameters and achieved the lowest validation loss.","The log-transform of the time-shift target could be reused whenever the quantity to regress spans several orders of magnitude and has a known sign; it converts a hard small-value regression into a well-conditioned one.","A natural next test is to tighten the constraint check with the error margin $\\zeta$ described in Section 2 and measure whether the network-only branch can then guarantee feasibility without any fallback; the paper explicitly leaves that analysis to future work.","Because the dataset is generated by the exact TSG, the approximation inherits the terminal-set and horizon assumptions of that solver; any deployment would need to re-run data generation for a different orbit or a different constraint cone angle."],"forward_implications":["A KAN-based TSG can run reference adjustment in about $0.03\\,\\mathrm{s}$ per step on a mid-range laptop CPU, which is fast enough for onboard real-time use in constrained rendezvous, a domain where the exact solver's worst case is about $0.85\\,\\mathrm{s}$.","Because CIKAN uses roughly 1.6 million parameters versus 209 million for the largest MLP-based CINN while achieving comparable or better training and validation loss, the memory footprint of onboard constraint enforcement can shrink dramatically.","The same hybrid pattern, predict with a network and fall back to an exact optimizer, transfers to other parameter-governor problems where the online optimization is the bottleneck.","Different neural models, including CIKAN, GRBF-CIKAN, RSW AF-CIKAN, and two MLP-based CINNs, all drive the deputy to the chief without violating the three constraints in the tested scenarios, so the reported benefit is not specific to one KAN basis choice."],"supporting_citations":[{"why":"Supplies the Kolmogorov-Arnold network architecture, the spline-based activation functions, and the approximation-theoretic discussion that motivate using KANs.","marker":"[Liu et al., 2024]"},{"why":"Develops the TSG for spacecraft rendezvous on elliptic orbits, which is both the exact solver being approximated and the source of the training dataset.","marker":"[Kim et al., 2024b]"},{"why":"Extends TSG to nonlinear cislunar dynamics and shows that TSG optimization with nonlinear models raises computational cost, the gap this paper targets.","marker":"[Kim et al., 2024a]"},{"why":"Introduces the TSG mechanism for spacecraft formation flying on circular orbits, establishing the time-shift reference adjustment that CIKAN learns.","marker":"[Frey et al., 2016]"},{"why":"Provides the parameter-governor framework of which the TSG is an add-on instance and which justifies the constraint-enforcement structure.","marker":"[Kolmanovsky and Sun, 2006]"},{"why":"Gives set-membership approximation theory for explicit MPC laws, referenced as the theoretical basis for conditions under which the approximate map πθ can be used.","marker":"[Canale et al., 2009]"},{"why":"Extends set-membership approximations of predictive control laws to the accuracy-complexity tradeoff, supporting the paper's use of a learned approximation.","marker":"[Canale et al., 2010]"},{"why":"States the representation theorem that underlies the KAN architecture's ability to represent continuous functions of many variables.","marker":"[Kolmogorov, 1957]"}],"fun_headline_variants":["KAN proxy for time-shift governor slashes fuel and compute","Hybrid KAN loop cuts rendezvous compute 62%, fuel 10%","Constraint-informed KAN matches TSG at lower cost","KAN-based governor enforces rendezvous constraints at lower cost","CIKAN beats MLP and exact TSG on elliptic orbits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The speed gain rests on the assumption that Algorithm 1 most often accepts the network prediction without falling back to the exact TSG solver, and that the one million training pairs sampled from the exact solver cover the states that appear in the Monte Carlo and extreme-case runs; if the fallback fires frequently or the dataset misses a region, the reported computation-time advantage and constraint enforcement could fail.","fun_headline_variants_meta":{"raw":{"variants":["KAN proxy for time-shift governor slashes fuel and compute","Hybrid KAN loop cuts rendezvous compute 62%, fuel 10%","Constraint-informed KAN matches TSG at lower cost","KAN-based governor enforces rendezvous constraints at lower cost","CIKAN beats MLP and exact TSG on elliptic orbits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000535,"raw_usage":{"total_tokens":2570,"prompt_tokens":945,"completion_tokens":1625,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":1535}},"tokens_in":561,"tokens_out":1625,"duration_ms":13333,"temperature":1.0,"reasoning_tokens":1535,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:09:45.050638+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Count the number of times Algorithm 1 calls the exact TSG solver over a large battery of Monte Carlo rendezvous runs, and separately disable the fallback and measure line-of-sight cone violations over states outside the training distribution; if the fallback rate is not close to zero, the 0.0329 second average mixes network and optimizer cost, and if violations appear without the fallback, the learned map does not itself enforce the constraints.","supporting_citations":[{"cited_title":"Time shift governor for coordinated control of two spacecraft formations","cited_arxiv_id":null,"evidence_quote":"Introduces the TSG mechanism for spacecraft formation flying on circular orbits, establishing the time-shift reference adjustment that CIKAN learns."},{"cited_title":"Parameter governors for discrete-time nonlinear systems with pointwise-in-time state and control constraints","cited_arxiv_id":null,"evidence_quote":"Provides the parameter-governor framework of which the TSG is an add-on instance and which justifies the constraint-enforcement structure."},{"cited_title":"Set membership approximation theory for fast implementation of model predictive control laws","cited_arxiv_id":null,"evidence_quote":"Gives set-membership approximation theory for explicit MPC laws, referenced as the theoretical basis for conditions under which the approximate map πθ can be used."},{"cited_title":"Set membership approximations of predictive control laws: the tradeoff between accuracy and complexity","cited_arxiv_id":null,"evidence_quote":"Extends set-membership approximations of predictive control laws to the accuracy-complexity tradeoff, supporting the paper's use of a learned approximation."},{"cited_title":"On the representation of continuous functions of many variables by superposition of continuous functions of one variable and addition","cited_arxiv_id":null,"evidence_quote":"States the representation theorem that underlies the KAN architecture's ability to represent continuous functions of many variables."}],"review_version":1}