{"id":"06b7b00a-e293-43f9-8a95-0f1025ec18c7","arxiv_id":"2509.11235","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"On a quadruple tank system, MPCs beat a SIMC-tuned decentralized PID for setpoint tracking with future setpoint preview, but the PID is competitive for disturbance rejection and without preview.","lead":"This paper compares three control algorithms, a decentralized PID, a linear MPC, and a nonlinear MPC, on a physical four-tank water system and in simulation. The MPCs tracked pre-announced setpoint changes far better than the PID, but the PID was competitive for disturbance rejection and when future setpoints were unavailable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The PID-vs-MPC conclusion is not yet robust: one PID tuning and one experimental run per controller leave the performance gap open to tuning or run artifacts.","rationale":"The paper is a transparent engineering comparison: the model identification is careful, the simulation studies are well structured, and the authors explicitly list limitations in Sec. 6.3.5. I do not see a mathematical flaw in the formulations. The central claim is also modest: the 'anticipatory advantage' of MPC is demonstrated by comparing MPC with and without future setpoint information (Simulations 1 vs 2), and this is robust to the noise-covariance concern because those simulations use R=diag(0.02) and sigma_a=1.0 for both plant and estimators. The main residual threat is fairness of the PID baseline: SIMC with a single Tc and no optimization, while MPC weights are tuned by trial-and-error to the same quadratic objective used for scoring. The paper concedes this, but the concession does not remove the risk. A tuning sweep in the simulation environment can settle it. I also note the single-run physical experiments as a secondary concern; repeated runs would strengthen but not invalidate the conclusion, since the simulations independently support the qualitative ranking. Therefore the reader's CONDITIONAL verdict is appropriate and no adjustment is needed.","tokens_in":28157,"tokens_out":11198,"duration_ms":144319,"concrete_test":"In the Simulation 1/2 environment, sweep the SIMC tuning parameter Tc over {10, 20, 50, 100, 200}, or directly optimize Kp, τi, τd against the NISE objective, and recompute Table 4. If the best-PID NISE falls below the reported LMPC/NMPC values for the with-preview case, the claim that MPCs outperform a well-tuned PID is tuning-dependent; if the gap persists across the sweep, the concern is largely settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that LMPC/NMPC 'perform better' than a decentralized PID for setpoint tracking rests on a comparison in which the PID is tuned once via SIMC with Tc=50 (Sec. 5.1.1), the MPC weights are set by trial-and-error (Secs. 5.2-5.3), and the physical evidence is one run per controller (Sec. 6.1.2). Section 6.3.5 explicitly concedes that the reported performance measures in Eq. (44) are the same quadratic tracking/input-rate objectives minimized by the MPCs. This is not an internal inconsistency, but it is a correctness risk: a different defensible PID tuning (or a non-representative run) could shrink or reverse the observed gap. The 'anticipatory advantage' component is better supported by Simulation 1 vs. Simulation 2, but those simulations inherit the same fixed PID tuning and MPC weights, so even that magnitude is tuning-dependent. The paper calls the PID well-tuned but provides no search over its tuning parameters.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper compares a decentralized SIMC-tuned PID, an LMPC, and an NMPC on a quadruple tank system, using both physical experiments and simulations. The QTS is modeled as a stochastic nonlinear continuous-discrete-time system with parameters estimated by maximum-likelihood prediction-error methods. The central empirical claim is that the LMPC and NMPC outperform the decentralized PID for tracking pre-announced time-varying setpoints, that this advantage essentially disappears when future setpoint information is withheld from the MPCs, and that for disturbance rejection the MPCs provide only marginal improvements over the PID. The NMPC yields slightly better tracking than the LMPC at the cost of higher input activity. The paper also provides a structured comparison of the three controllers under identical setpoint sequences and discusses the tuning dependence of its conclusions.","tokens_in":28511,"tokens_out":6430,"duration_ms":74513,"significance":"If the claim holds, the paper provides a useful and fairly comprehensive benchmark: on a standard QTS, the measurable advantage of MPC over a well-tuned decentralized PID is primarily anticipatory, not regulatory. The study is valuable for its systematic experimental protocol, the use of the same identified model as the basis for all three controller designs, and the explicit acknowledgment of the limitations of the comparison. However, the strength of the central claim is currently limited by the use of one experimental run per controller, a single PID tuning, a manually manipulated measurement-noise covariance in the identification, and an ambiguity about the parameters used in the simulation plant.","major_comments":[{"comment":"The experimental comparison is based on one run per controller (Sec. 6.1.2). The headline NISE values in Table 4 (PID 9.063, LMPC 1.637, NMPC 1.423) are point estimates from single trajectories. Given the stochastic dynamics in (1) and the process disturbances, these differences could lie within run-to-run variability. The paper should provide repeated experiments or Monte Carlo simulations using the identified stochastic model, with confidence intervals for NISE, NIAE, and NISΔU. Without such an uncertainty analysis, the central claim rests on a single realization.","section":"Sec. 6.1.2, Tables 4-5"},{"comment":"The PID is tuned once with SIMC using Tc=50, while the MPC weights (Q=10I, S=I) and horizon N=160 are chosen by trial-and-error. Section 6.3.5 itself concedes that different tunings could change the results and that the performance measures in Eq. (44) mirror the MPC cost. This makes the reported performance gap a property of the chosen tuning configurations rather than of the algorithms as such. The authors should add a sensitivity analysis over Tc and Q/S, or use optimization-based tuning of both designs under the same objective, before claiming that the MPCs 'perform better' than the PID. This is load-bearing for the paper's main conclusion.","section":"Secs. 5.1.1, 5.2, 5.3, 6.3.5"},{"comment":"The measurement noise covariance for the upper tanks is manually inflated by a factor 1000 (Sec. 4.2.2), and the subsequent ML-PEM estimation yields r_3^2 = r_4^2 = 1e-5, effectively giving zero weight to the upper-tank measurements. The estimated parameters A3, A4, and gamma1 differ substantially from the nominal values (Table 1). Since the same identified model is used to tune the PID and to design the MPCs and their state estimators, the controller comparison may be influenced by this ad hoc identification choice. Please validate the noise covariance on an independent steady-state data set, or demonstrate that the conclusions are invariant to a range of plausible R values.","section":"Sec. 4.2.2, Table 3"},{"comment":"The text states that 'we apply the nominal parameters in Table 1 for all four simulation studies', while the controllers were designed using the estimated parameters in the same table. This plant-model mismatch is not discussed, and the claim in Sec. 6.2.1 that the CD-KF and CD-EKF have 'perfect knowledge about the systems' is misleading if the plant uses the nominal parameters. Specify explicitly which parameter set is used for the simulated plant and which is used in the controllers and estimators, and discuss the implications of any mismatch for the interpretation of Simulations 1-4.","section":"Sec. 6.2"}],"minor_comments":[{"comment":"The units of r^2 are given as [m^2], but the measurement equation (1b) uses y in cm; the covariance units should be reconciled.","section":"Table 3"},{"comment":"The measures are called 'normalized' NISE/NIAE, but the definitions are simple averages of squared/absolute errors; clarify the normalization or rename the quantities.","section":"Eq. (44)"},{"comment":"The factor 1000 used to inflate the upper-tank measurement variances is introduced with only a qualitative turbulence argument; provide a quantitative justification or a sensitivity check.","section":"Sec. 4.2.2"},{"comment":"The abstract describes the PID as 'well-tuned'; given the single SIMC tuning and the paper's own caveats, consider wording such as 'SIMC-tuned' to avoid overstatement.","section":"Abstract / Sec. 5.1.1"},{"comment":"There are typographical issues (e.g., 'di fferent' in several places, 'Futhermore' in Sec. 6.3.4, 'tutotorial' in the Pannocchia reference). A careful proofread is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and presents a useful experimental benchmark. The main concern is not novelty but robustness of the central comparison: one run per controller, one PID tuning, and an ad hoc noise-inflation step in the identification. These are fixable with additional sensitivity analyses, repeated experiments or Monte Carlo studies, and clarifying the simulation plant-model setup. I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a competent, well-scoped comparison of decentralized PID, LMPC, and NMPC on a physical quadruple tank system, and its central claim is credible. The MPCs beat the PID for tracking pre-announced setpoints, but the edge is mostly anticipatory—Simulation 2 (Table 4) shows the PID actually has lower NISE and NIAE than both MPCs when future setpoint info is removed. Disturbance rejection gains are marginal. That qualitative pattern is likely robust.\n\nThe paper extends the authors' earlier conference work with ML-PEM identification of a stochastic continuous-discrete model, disturbance-augmented CD-KF/CD-EKF estimation, and simulation studies that isolate future-setpoint information. The modeling and estimator details are careful, the experimental implementation is real, and the paper is unusually honest: Section 6.3.5 explicitly acknowledges the tuning asymmetry and that the performance metrics align with the MPC objective.\n\nThe soft spots are real but not fatal. One experimental run per controller and a single SIMC PID tuning mean the numbers in Tables 4-5 could shift with different tuning or a different run; the stress-test note is right about that, and the paper's own caveat covers it. The simulation plant-model setup is ambiguous: Section 6.2 says 'nominal parameters in Table 1' while the controllers use estimated parameters—if the simulated plant is the nominal model, the MPCs are evaluated under a plant-model mismatch not present in the experiments. That needs clarification. The manual inflation of the measurement noise covariance (factor 1000, Section 4.2.2) and the near-zero upper-tank noise estimates (Table 3) are a bit hacky; they could bias parameter estimates, though the validation GOF is decent.\n\nThis paper is for control practitioners and educators who want a concrete benchmark demonstration of when MPC pays off. It is not a theoretical contribution, and the central numbers are not definitive, but the qualitative conclusions are sound and the limitations are honestly stated. It deserves a serious referee; I would recommend major revision to tighten the simulation setup and add a second experimental run or at least a sensitivity check on PID tuning.","headline":"A solid, honest engineering comparison; the key claim that MPC's advantage over a SIMC-tuned PID is mostly anticipatory holds up, but the numeric magnitudes are tuning-dependent.","tokens_in":28956,"tokens_out":4856,"would_cite":true,"duration_ms":53990,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MPC’s advantage over a tuned PID lies in anticipating setpoints, not in rejecting disturbances.","keywords":["Quadruple tank system","Decentralized PID","Linear model predictive control","Nonlinear model predictive control","ML-PEM parameter estimation","SIMC tuning","Continuous-discrete Kalman filter","Anticipatory control"],"falsifier":"On the physical quadruple tank, run the same setpoint sequence with the LMPC and NMPC given only current setpoints (setpoint held constant over the prediction horizon); if either MPC still achieves lower NISE/NIAE than the SIMC-tuned PID, the claim that the primary MPC advantage is anticipatory is falsified.","tokens_in":28133,"feed_emoji":"🎛️","tokens_out":4637,"duration_ms":51629,"temperature":0.7,"pith_summary":"This paper compares three control strategies on a physical quadruple tank system and in simulation: a decentralized PID tuned by SIMC rules, a linear MPC, and a nonlinear MPC. It finds that the MPCs track pre-announced time-varying setpoints far better than the PID, but only slightly better for disturbance rejection. The central claim is that the MPC advantage is mainly anticipatory: when future setpoint information is removed and only current setpoints are given, the PID actually achieves better tracking errors in simulation. The paper matters because it quantifies when the extra complexity of MPC pays off in a realistic benchmark.","feed_headline":"MPC beats PID by seeing future setpoints, not by better feedback","feed_subtitle":"On a four-tank rig, MPC wins on announced setpoint changes but barely improves disturbance rejection over PID.","key_machinery":"The comparison is carried by a common stochastic continuous-discrete-time model of the four-tank process, identified with a maximum-likelihood prediction-error method (ML-PEM). The NMPC uses the nonlinear model in its optimal control problem with a continuous-discrete extended Kalman filter (CD-EKF) for state and disturbance estimation; the LMPC uses the same model linearized, with a continuous-discrete Kalman filter (CD-KF). The PID is tuned from transfer functions of the same linearized model using SIMC rules. The decisive mechanism is the MPC prediction horizon: with a 13-minute horizon and pre-announced setpoints, the optimizer can move valves before the setpoint change arrives, which a","core_discovery":"The paper claims that on the quadruple tank system, the performance gap between model predictive control and a well-tuned decentralized PID is driven by the MPC's ability to incorporate future setpoint information, not by superior feedback disturbance rejection. Experimentally, LMPC and NMPC achieve markedly lower tracking-error norms (NISE about 1.6 and 1.4 vs PID's 9.1) and much lower input movement (NISΔU about 12–29 vs PID's 249) on a preset setpoint sequence. In simulation, when the MPCs are only given current setpoints, the PID's tracking errors become smaller than the MPCs'. For large deterministic disturbances the MPCs are only slightly better, and for stochastic disturbances they ar","pith_inferences":["The paper's logic suggests a testable general rule: for any well-modeled process, the tracking advantage of MPC over tuned PID scales with the amount and reliability of future setpoint information, not with model complexity; an optimization-based PID tuned on the same objective as the MPC would be a sharper baseline.","Because the simulation studies assume the controller and estimator know the true noise covariances, the 'no-future-information' result is an upper bound on PID competitiveness; with model mismatch, the PID's integral action may look even better, so the claim deserves a robustness test with perturbed plant parameters.","The manual 1000x inflation of upper-tank measurement noise variances in Section 4.2.2 is a fragile identification step; re-estimating the model with proper noise modeling (e.g., estimating separate sensor variances without inflation) and re-running the comparison would show whether the MPC advantage is sensitive to the identification procedure."],"forward_implications":["In industrial loops where setpoint changes are known in advance (batch transitions, grade changes, scheduled trajectories), MPC can deliver large reductions in both tracking error and valve wear compared with a well-tuned decentralized PID.","When disturbances are unmeasured and no future information exists, a SIMC-tuned decentralized PID with integral action is competitive with MPC; spending on MPC for pure regulatory control may yield little tracking benefit.","NMPC vs LMPC: nonlinearity pays only a small tracking dividend on this rig in the tested operating range, while roughly doubling the input movement; linear MPC may be the cost-effective choice for mildly nonlinear tank systems.","The large reduction in input rate of movement (NISΔU) with MPCs suggests that MPC can extend actuator life and reduce wear even when tracking gains over PID are modest."],"fun_headline_variants":["MPC's edge over PID comes from future setpoints, not feedback","Future setpoints, not feedback, explain MPC's win over PID","On four-tank rig, MPC beats PID by knowing tomorrow's setpoints","MPC outperforms PID mainly when setpoints are pre-announced","Why MPC beats PID on tanks: it peeks at future setpoints"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The comparison's conclusions rest on the identified stochastic model being accurate enough that the simulated no-future-setpoint case faithfully represents real closed-loop behavior, yet that model depends on a manually inflated measurement noise covariance (factor 1000) whose correctness is untested.","fun_headline_variants_meta":{"raw":{"variants":["MPC's edge over PID comes from future setpoints, not feedback","Future setpoints, not feedback, explain MPC's win over PID","On four-tank rig, MPC beats PID by knowing tomorrow's setpoints","MPC outperforms PID mainly when setpoints are pre-announced","Why MPC beats PID on tanks: it peeks at future setpoints"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000161,"raw_usage":{"total_tokens":1124,"prompt_tokens":846,"completion_tokens":278,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":182}},"tokens_in":590,"tokens_out":278,"duration_ms":3274,"temperature":1.0,"reasoning_tokens":182,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:50:37.921795+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the physical quadruple tank, run the same setpoint sequence with the LMPC and NMPC given only current setpoints (setpoint held constant over the prediction horizon); if either MPC still achieves lower NISE/NIAE than the SIMC-tuned PID, the claim that the primary MPC advantage is anticipatory is falsified.","supporting_citations":[],"review_version":1}