{"id":"2afbe77a-7c50-4666-b7c7-504dd23dea70","arxiv_id":"2511.13588","paper_version":2,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A nonparametric policy constructed from offline MPC solutions accelerates control by orders of magnitude while preserving recursive feasibility and bounded optimality gap under sufficient data coverage.","lead":"The paper introduces a data-driven nonparametric policy that replaces online optimization in Model Predictive Control with a fast lookup rule built from offline solutions. This approach trades some optimality for speed while providing guarantees on feasibility and performance under data coverage conditions.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Coverage conditions may not ensure positive invariance of the data-covered set under the greedy policy, risking invalidation of the upper bound during online execution.","rationale":"The reader's weakest assumption directly identifies the coverage condition needed for the upper bound to hold on encountered states. The concern above sharpens that point by noting the missing invariance argument that would make the condition sufficient for closed-loop execution. This does not reject the overall approach but keeps the verdict conditional until the dynamic coverage property is verified or strengthened.","tokens_in":1656,"tokens_out":372,"duration_ms":37207,"concrete_test":"From the analysis section, extract the precise definition of the coverage condition (e.g., any ball or set around data points) and the policy rule; for a simple double-integrator or linear system with known optimal MPC, generate a modest offline dataset satisfying the coverage assumption, apply the greedy policy from a covered initial state, and check whether every subsequent state remains inside the covered set. If any trajectory exits, the invariance fails and the recursive-feasibility claim requires an additional assumption or terminal-set construction.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that the offline data's sufficient coverage makes the constructed upper bound on the cost-to-go valid for every state visited by the closed-loop system. This underpins both recursive feasibility and the bounded optimality gap. However, coverage is a static property of the dataset; the analysis does not appear to establish that the nonparametric greedy policy (lookup rule w.r.t. the upper bound) maps the covered set into itself. If a trajectory leaves the covered region, the bound ceases to be valid and the guarantees collapse. This is the least secure link because the paper's trade-off between data volume and bound tightness assumes the dynamic trajectory never escapes the static cover.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes a data-driven framework to accelerate MPC by constructing a nonparametric greedy policy from offline MPC solutions. The policy is defined with respect to a constructed upper bound on the optimal cost-to-go and can be implemented as a fast lookup rule. Under sufficient coverage conditions on the offline data, the analysis claims recursive feasibility and a provable bounded optimality gap, with an explicit trade-off between data volume and bound tightness. New solutions can be added without retraining, and experiments report 100-1000x speedup over standard MPC with modest optimality loss.","tokens_in":1806,"tokens_out":437,"duration_ms":39820,"significance":"If the central guarantees hold, the work offers a meaningful contribution to real-time control by providing a fast, data-driven MPC approximation with explicit performance bounds and continual improvement capability. The nonparametric nature and lack of retraining requirement are practical strengths that could enable deployment in low-latency applications where traditional MPC is prohibitive.","major_comments":[{"comment":"Analysis section (as described in the abstract): the claim that sufficient coverage conditions ensure recursive feasibility and bounded optimality gap is load-bearing, yet the manuscript does not establish that the greedy policy (nonparametric lookup w.r.t. the upper bound) renders the data-covered set positively invariant. Coverage is a static property of the offline dataset; without a proof that closed-loop trajectories remain inside this set, the upper bound may cease to be valid online, collapsing both feasibility and the gap guarantee. This directly affects the stated trade-off between data volume and bound tightness.","section":"Analysis section"}],"minor_comments":[{"comment":"The abstract refers to 'sufficient coverage conditions' without a forward reference to their precise definition or the theorem that invokes them; adding this would improve readability.","section":"Abstract"},{"comment":"Experiments section: while speedup factors are reported, including the state/input dimensions of the benchmark problems and the hardware platform would allow better assessment of the practical speedup.","section":"Experiments"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"Thank you for the careful review and constructive feedback on our manuscript. We address the major comment regarding the analysis of recursive feasibility and the positive invariance of the covered set under the greedy policy. We agree that this aspect requires strengthening and outline the planned revisions below.","responses":[{"response":"We thank the referee for highlighting this critical point. We acknowledge that while the manuscript states the guarantees under sufficient coverage conditions, the explicit proof establishing that the greedy policy renders the data-covered set positively invariant is not developed in sufficient detail. This leaves the online validity of the upper bound and the resulting recursive feasibility and bounded optimality gap insufficiently supported. In the revised manuscript, we will expand the Analysis section with a dedicated lemma proving positive invariance. The proof will show that, for any state in the covered set, the control selected by the nonparametric greedy policy (with respect to the constructed upper bound) produces a successor state that remains covered. This follows from the consistency of the upper bound with the system dynamics and the fact that the offline dataset includes complete optimal trajectories, ensuring the coverage condition propagates along closed-loop trajectories. We will also clarify how this invariance underpins the explicit trade-off between data volume and bound tightness. These additions will be placed immediately following the coverage definition to make the argument self-contained.","revision_made":"yes","referee_comment":"[Analysis section] Analysis section (as described in the abstract): the claim that sufficient coverage conditions ensure recursive feasibility and bounded optimality gap is load-bearing, yet the manuscript does not establish that the greedy policy (nonparametric lookup w.r.t. the upper bound) renders the data-covered set positively invariant. Coverage is a static property of the offline dataset; without a proof that closed-loop trajectories remain inside this set, the upper bound may cease to be valid online, collapsing both feasibility and the gap guarantee. This directly affects the stated trade-off between data volume and bound tightness."}],"tokens_in":1275,"tokens_out":408,"duration_ms":48013,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main thing here is a nonparametric policy built from offline MPC runs that picks actions greedily against a data-derived upper bound on cost-to-go. Under coverage conditions the authors claim recursive feasibility and a bounded optimality gap, plus an explicit link between data volume and bound tightness. New solutions can be added without retraining.","headline":"This gives a nonparametric lookup policy for fast MPC with coverage-based feasibility and gap bounds, but the policy may not keep trajectories inside the covered set.","tokens_in":2284,"tokens_out":138,"would_cite":false,"duration_ms":53937,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[],"headline":"Nonparametric greedy policy on Lipschitz cost-to-go upper bounds for MPC acceleration","alignment":"orthogonal","rationale":"The paper's core construction (Definition 1 nonparametric policy π_D via arg min of J_i + λ‖x−x_i‖, J_λ_ub upper bound, coverage conditions in Theorem 2, erosion X^{-ε} for recursive feasibility in Propositions 2-3) is a standard approximate-MPC / nearest-neighbor control scheme under Lipschitz and coverage assumptions. It neither invokes nor parallels any RS forcing chain element (J-cost functional equation, φ-ladder, 8-tick periodicity, Alexander-duality D=3, or reality_from_one_distinction). Domain is applied control theory; RS has no opinion on receding-horizon policies or data-coverage PAC bounds.","tokens_in":54177,"confidence":"high","tokens_out":185,"duration_ms":13369,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A nonparametric policy from offline MPC solutions replaces online optimization while guaranteeing recursive feasibility and a bounded optimality gap.","keywords":["data-driven MPC","model predictive control","nonparametric policy","recursive feasibility","optimality gap","offline data coverage","real-time control","cost-to-go bound"],"falsifier":"Run the policy on a trajectory that reaches a state outside the covered region and observe whether feasibility is lost or the realized optimality gap exceeds the predicted bound.","tokens_in":2563,"feed_emoji":"🚀","tokens_out":662,"duration_ms":33162,"temperature":0.7,"pith_summary":"This paper develops a data-driven way to make model predictive control fast enough for real-time use. It replaces the slow online solver with a simple lookup rule built from previously computed MPC solutions. The rule picks the action that looks best according to an upper bound on the remaining cost, and the analysis shows that if the offline data covers the states well enough, the resulting closed-loop behavior stays feasible at every step and loses only a limited amount of performance compared with true optimal MPC. The bound on the loss gets tighter as more offline solutions are added, giving a direct link between data volume and guarantee strength. Because new solutions can be inserted directly into the lookup table, the policy can keep improving without any retraining step.","feed_headline":"Data-driven policy speeds up MPC by 100-1000x with provable bound","feed_subtitle":"Offline solutions create a fast lookup rule that stays feasible and loses only a limited amount of optimality under coverage conditions.","key_machinery":"The nonparametric greedy policy that selects the control minimizing immediate cost plus the data-derived upper bound on cost-to-go.","core_discovery":"The central claim is that a greedy policy with respect to a constructed upper bound on the optimal cost-to-go, formed from a finite set of offline MPC solutions, can be substituted for repeated online optimization. When the offline data satisfies sufficient coverage conditions, this policy remains recursively feasible and its infinite-horizon cost lies within an explicitly characterized gap of the optimal MPC cost, with the gap shrinking as the data set grows.","pith_inferences":["The coverage condition could be monitored online so that new data is collected only when the current bound risks becoming invalid.","The same construction might accelerate other receding-horizon planners beyond classical MPC.","Because the policy is a simple lookup, it could run on embedded hardware that lacks the resources for real-time optimization."],"forward_implications":["The policy executes 100 to 1000 times faster than solving MPC online.","Adding any new MPC solution immediately tightens the bound without requiring retraining.","An explicit trade-off exists between the number of offline samples and the size of the optimality gap.","The same policy structure applies to any real-time control task where standard MPC is currently too slow."],"fun_headline_variants":["Offline MPC data constructs greedy policy with recursive feasibility guarantees","Greedy policy from offline solutions replaces online optimization with bounded gap","Finite MPC data yields nonparametric lookup maintaining feasibility and bounds","Data coverage enables greedy cost-to-go policy with explicit optimality gap"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The offline data set must be dense enough that the constructed upper bound on cost-to-go stays valid for every state the system actually visits during online operation.","fun_headline_variants_meta":{"raw":{"variants":["Offline MPC data constructs greedy policy with recursive feasibility guarantees","Greedy policy from offline solutions replaces online optimization with bounded gap","Finite MPC data yields nonparametric lookup maintaining feasibility and bounds","Data coverage enables greedy cost-to-go policy with explicit optimality gap"]},"model":"grok-4.3","cost_usd":0.01049,"raw_usage":{"total_tokens":4531,"prompt_tokens":616,"num_sources_used":0,"completion_tokens":65,"cost_in_usd_ticks":104903000,"prompt_tokens_details":{"text_tokens":616,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3850,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":616,"tokens_out":65,"duration_ms":49059,"temperature":1.0,"reasoning_tokens":3850,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-21T18:17:16.857295+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Run the policy on a trajectory that reaches a state outside the covered region and observe whether feasibility is lost or the realized optimality gap exceeds the predicted bound.","supporting_citations":[],"review_version":1}