Many industrial feedback loops pair a PID controller with a fuzzy supervisor. The PID part is the familiar controller that reacts to the current error, its accumulated history, and its rate of change. The fuzzy supervisor is a table of hand-readable IF-THEN rules, such as "if the error is large and growing, raise the proportional gain," and it retunes the PID gains as conditions change. These hybrids clearly outperform fixed gains. On a 2-DOF helicopter, fuzzy-PID cut pitch rise time from 7.5 s to 1 s and overshoot from 40% to 6% (Huynh et al.). The rule table, however, grows exponentially with the number of inputs, and it is tuned by swarm or evolutionary search that comes with no guarantee of any kind (Liu et al., Li et al.). This proposal asks a narrow, answerable question. Given a good but bloated fuzzy-PID controller, can we pick a small subset of its rules with a provable guarantee on how faithfully the small controller reproduces the large one where it matters, and get the full trade-off between rule count and fidelity from one cheap computation?
The idea is to treat rule selection as a coverage problem rather than as direct optimization of closed-loop cost. Closed-loop cost has no diminishing-returns structure that greedy selection can exploit. Coverage does. We record the error states the plant actually visits under the large controller. We then choose $k$ rules so that every visited state is close to some kept rule, with closeness measured in both state and gain, and with each state weighted by how sensitive closed-loop performance is to gain errors there. This objective is a facility-location function, and the simple greedy algorithm provably reaches at least 63% of its optimum. Saving each greedy prefix then traces the whole rules-versus-fidelity curve in a single pass.
Core Claim
Fuzzy rule-base compression can be posed as weighted facility location over visited operating states, which makes it submodular. That structure buys three properties that metaheuristic tuning lacks. Selection needs no closed-loop rollouts beyond one data-collection phase. The result carries an approximation certificate on a quantity that upper-bounds the gain-schedule error. One greedy run yields the whole Pareto curve across all rule budgets $k$. The certificate is only as useful as its link to closed-loop behavior, so the key new ingredient is a sensitivity weighting that ties coverage error to closed-loop cost and to stability margin. Plain clustering of visited states has no such weighting.
How This Differs From Existing Work
The file reports no prior work combining fuzzy-PID with submodular greedy guarantees, so the comparison is with adjacent work.
The construction is borrowed from CRAIG. That paper bounds the gap between subset and full-data gradients with a facility-location objective, which greedy maximizes to within $1-1/e$. We copy that construction but replace "data points whose gradients must be matched" with "rules whose gain outputs must be matched along trajectories."
The curve-from-prefixes idea comes from ParetoGreedy in Vombatkere and Terzi, which recovers an approximate utility–cost Pareto frontier from greedy prefixes. Their setting is static graphs and recommender data. Ours uses rule count as the cost, a dynamical system generates the ground set, and the utility has to be validated in closed loop.
MP-MoE motivates the framing. It shows that picking modular components (mixture-of-experts experts) by greedy diversity-aware selection beats picking the individually highest scorers. Fuzzy rules are a close analogue of experts, but MP-MoE never touches feedback loops.
The tuning-without-guarantees problem is set by Liu et al., Li et al., and the AGE-MOEA-II work of Multi-Objective F-FOPID. We do not replace their tuning of the large controller; we take that controller as the teacher.
Sharma et al. constrains the design. They obtain global optimality for continuous PID gains via gradient dominance, but only for linear plants, and their model-free proof does not extend to the derivative gain. That is why we do not try to optimize closed-loop cost directly. Instead we compress a teacher that already handles nonlinearity and derivative scheduling.
Research Program
The teacher is a dense fuzzy-PID supervisor: for example, a 7×7 grid over error and error-rate, giving 49 rules, each outputting a gain correction. We roll it out across a randomized set of references and disturbances and collect visited states $x_i = (e_i, \Delta e_i)$ together with the teacher's gain vector $K(x_i)$. The candidate ground set $V$ is the teacher's rule centers. Optionally it is enlarged with extra centers placed at visited states, which addresses the static-ground-set limitation by building the ground set from the trajectories.
The objective is
$$f(S) = \sum_i w_i \max_{j \in S} \big( C - d(x_i, x_j) \big),$$
where $S$ is the set of kept rules and $C$ is a constant larger than every distance $d$. The distance is $d(x_i, x_j) = \|x_i - x_j\| + \beta \|K(x_i) - K(x_j)\|$, with $\beta$ trading state distance against gain distance. The weight $w_i$ is the finite-difference sensitivity of the closed-loop cost (ITAE, the time-weighted absolute error) to a small gain perturbation applied when the plant is near $x_i$, estimated from a few perturbed rollouts.
The function $f$ is monotone submodular. Maximizing it is equivalent to minimizing the weighted nearest-rule distance. Under a Lipschitz assumption on the teacher's schedule, that distance upper-bounds the weighted gain error of a student that copies the nearest kept rule's consequent. We will also test membership-weighted interpolation among kept rules, but the bound covers only the nearest-rule student.
Greedy selection gives the $1-1/e$ guarantee from Kempe and Tardos. We will also report the data-dependent bound obtained by summing the top-$k$ marginal gains at each step, which in practice is often far tighter than 63%.
On the theory side, we will try to state a sufficient condition linking the two worlds. If the student's gain error on the visited set stays below the teacher's stability margin to gain perturbation, closed-loop stability is preserved locally around the teacher's trajectories. This responds to the file's Gap 3: a solution that is 63% of optimal on some objective can still destabilize a plant. We expect the result to be conservative and local. We will measure how often the empirical student is stable even when the condition fails, and we will not claim more than that.
Baselines
The do-nothing reference is the fixed-gain PID used as the starting point in the cited studies. It anchors how much of the hybrid's advantage survives compression.
The upper reference is the uncompressed teacher. On the 4-DOF manipulator, it is a dual-fuzzy PSO-tuned controller following Liu et al., which reports positioning time reduced from 2.8 s to 1.8 s and positioning error from 7.6×10⁻³ rad to 2.8×10⁻⁵ rad. On the helicopter, it follows Huynh et al., which reports 1 s rise time and 6% overshoot. These are their published figures. Our reimplementations will be re-measured, and the comparison is against our own teacher under identical simulations.
The strongest competitor is direct metaheuristic subset search. A PSO, or a genetic algorithm with binary masks, searches over $k$-rule subsets and scores each candidate by closed-loop ITAE. It optimizes the true target rather than a surrogate. We run it once per $k$ and give it a total simulation budget equal to our data-collection plus sensitivity-estimation rollouts. We also report it at 10× that budget. Beating it means matching its ITAE within its seed-to-seed variance at equal budget while also delivering the whole $k$-curve and a certificate.
Two simpler compressions are also included. One is a coarser uniform grid, such as 3×3 membership functions. The other is k-means or k-medoids on visited states, with rule consequents fitted by least squares.
The decisive ablation runs the identical greedy procedure with uniform weights $w_i = 1$ and $\beta = 0$. This isolates whether the sensitivity weighting and the gain term, rather than coverage alone, carry the benefit.
Why not just…?
Why not just run PSO over rule subsets on closed-loop cost? This is the real threat. With enough rollouts it may find better subsets than any surrogate, because it optimizes the quantity we care about. Its weaknesses are documented in the papers that use metaheuristics. Li et al. needed Cauchy mutation and opposition-based learning to escape the premature convergence of standard GOA. Liu et al. flag online PSO's computational cost as a risk to real-time use. It also must be rerun for every $k$. The equal-budget and 10×-budget comparison settles the question. If PSO wins clearly at equal budget, our contribution shrinks to the certificate and the free Pareto curve, and we will say so.
Why not just cluster visited states with k-means? K-medoids on visited states is nearly the unweighted version of our objective. If it matches us, the weighting idea has failed. Without sensitivity weights it spends rules on frequently visited, low-consequence regions such as the near-zero-error steady state, and too few on rare high-consequence transients. The uniform-weight ablation and k-medoids baseline test exactly this. We will report the result whichever way it falls.
Why not just use fewer membership functions, or tune a smooth gain schedule by policy gradient? A coarser grid is the simplest compression, but it places rules uniformly regardless of where the plant operates. Policy gradient carries optimality guarantees only in the linear, PI-only regime of Sharma et al., so it does not cover nonlinear plants with derivative scheduling. Both appear as baselines. Policy-gradient PID is evaluated on the Controlgym chemical reactor, where its assumptions hold.
Proposed Experiments
The plants are four testbeds from the file. Three are simulated nonlinear plants: the Quanser-style 2-DOF helicopter, modeled with Euler–Lagrange equations including cross-coupling; the 4-DOF manipulator with Coriolis and gravity terms; and PMSM speed control, where the traditional fuzzy PI shows about 10.3% start-up overshoot in Li et al.. The fourth is the 8-dimensional chemical reactor in Controlgym, used for the comparison against policy-gradient PID.
Metrics are rise time, overshoot, settling time, ITAE, the fraction of randomized disturbance episodes that remain stable, rule evaluations per control step, and measured cycle time on an embedded-class processor. Every comparison is evaluated on held-out references and disturbances not used during selection.
The first decisive check is the shape of the trade-off curve. Plotting student ITAE against $k$ for greedy-weighted, greedy-uniform, k-medoids, coarse grid, and PSO should show the greedy-weighted method reaching within 5% of teacher ITAE at substantially fewer rules than the uniform-weight and grid variants.
The second check tests whether the certificate predicts anything. Across $k$ and plants, we will test whether the certified coverage gap correlates with the closed-loop ITAE gap and with stability loss. A certificate that does not track closed-loop behavior would be mathematically correct and practically empty.
The third check is robustness under distribution shift: plant-parameter changes such as manipulator payload and PMSM load torque that move the visited states. Here the visited-state ground set is most likely to fail. We will measure how much re-collection restores performance.
For the PMSM loop we will report whether the compressed supervisor fits within the cycle budget that rules out online PSO. The file contains no timing number for this; the 10–20 kHz figure is a typical current-loop rate, not a reported baseline result.
References
- Implementation of Fuzzy-PID Controller for 2-DOF Helicopter
- Self-Tuning Control of Manipulator Positioning Based on Fuzzy PID and PSO Algorithm
- Improved GOA-based fuzzy PI speed control of PMSM with predictive current regulation
- Coresets for Data-efficient Training of Machine Learning Models
- Computing Approximate Pareto Frontiers for Submodular Utility and Cost Tradeoffs
- Breaking the Echo Chamber: A Dynamic Ensemble Pruning Perspective on MoE
- Multi-Objective Optimized Fuzzy Fractional-Order PID Control for Frequency Regulation in Hydro–Wind–Solar–Storage Systems
- Globally Optimal Policy Gradient Algorithms for Reinforcement Learning with PID Control Policies
- Maximizing the spread of influence through a social network
- Controlgym: Large-scale control environments for benchmarking reinforcement learning algorithms