Unverified 2026

Cost-Map Finite-Action Head

Implementation & benchmark of arXiv:2608.22416 — A Simple and Extremely Efficient Predictive Control for Power Converters

Usefulness6/10

Difficulty5/10

Novelty6/10

Architecture Speedup Stability Rl Optimizer World model Inference speedup Control theory Optimization Linear algebra Geometry

Source paper: A Simple and Extremely Efficient Predictive Control for Power Converters arXiv:2608.22416 ⓘ · analyzed Sep 1, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Replace online enumeration over a finite action set with a classifier or lookup map whose regions directly return the action minimizing a one-step predictive-control cost. For affine dynamics and quadratic tracking loss, exact action regions are separated by pairwise cost boundaries, so the approximation can be audited against exhaustive predictive control rather than treated as an unconstrained policy.

Formulas

$$x_i^+=Ax+Bu_i+c,\qquad J_i(x)=(x_i^+-r)^TQ(x_i^+-r)+u_i^TRu_i,\qquad \pi(x)=\arg\min_{i\in\{1,\ldots,M\}}J_i(x).$$

$$J_i(x)-J_j(x)=2x^TA^TQB(u_i-u_j)+d_{ij},$$ $$d_{ij}=(Bu_i+c-r)^TQ(Bu_i+c-r)+u_i^TRu_i-(Bu_j+c-r)^TQ(Bu_j+c-r)-u_j^TRu_j.$$

$$\mathcal{R}_i=\{x:J_i(x)\le J_j(x)\ \forall j\},\qquad \operatorname{margin}(x)=J_{(2)}(x)-J_{(1)}(x),$$ $$\text{where }J_{(1)}\le J_{(2)}\text{ are the smallest and second-smallest action costs.}$$

Mathematical statement

Let the predicted state under action $u_i$ be $x_i^+=Ax+Bu_i+c$, where $x\in\mathbb{R}^n$, $u_i$ is one of $M$ discrete actions, $A,B,c$ are dynamics parameters, and $r$ is the target. Define the one-step cost $J_i(x)=(x_i^+-r)^TQ(x_i^+-r)+u_i^TRu_i$, with positive-semidefinite matrices $Q$ and $R$. The exact action map is $\pi(x)=\arg\min_iJ_i(x)$. The boundary between actions $i$ and $j$ is $J_i(x)-J_j(x)=0$; because $A$ and $c$ are shared across actions, the quadratic state term cancels and this boundary is affine. A neural classifier $g_\phi(x,r)$ learns the resulting winner regions, while a margin monitor falls back to exact enumeration near boundaries.

Implementation notes

(1) Integration point: use this as the discrete action-selection head of a model-based RL agent, neural controller, or learned-dynamics policy. Keep exhaustive one-step MPC as an oracle and safety fallback. The controller supplies the current state $x$ and target $r$; the head returns one of the $M$ discrete actions. (2) Pseudocode: sample states and targets; compute $x_i^+=Ax+Bu_i+c$ and every $J_i$; assign label $y=\arg\min_iJ_i$ and margin $m=J_{(2)}-J_{(1)}$; train $g_\phi(x,r)$ with cross-entropy, oversampling or weighting small-margin examples. At inference, compute the classifier action and confidence. If confidence is above a threshold and the predicted normalized margin is above a threshold, execute it; otherwise evaluate all $J_i$ and execute the exact argmin. A cheaper version precomputes the affine pairwise boundary coefficients and uses a lookup table. (3) Computed from the mechanism: exact labels, costs, margins, and analytic boundaries. Estimated empirically: classifier confidence calibration, fallback threshold, latency, and closed-loop cost. (4) First cheap experiment: use a two-dimensional linear system with eight converter-like actions, train a two-layer MLP on 100,000 sampled states, and compare exhaustive argmin, classifier-only, and classifier-plus-fallback controllers. Measure action disagreement, quadratic tracking cost, boundary localization error, and CPU time. The quantitative prediction is that disagreements concentrate near $m=0$; a fallback restricted to low-margin states should reduce wrong actions below 1% while eliminating at least 80% of cost evaluations. Analytic and learned decision boundaries should agree within 10% relative localization error, and closed-loop cost should match exhaustive MPC away from the boundary.

Verification

This idea has not been verified yet.

Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.

Artifacts

Artifacts unavailable.