What problem does it solve?
This skill helps you tune hyperparameters or system configuration by turning a discrete search problem with measurable objectives and constraints into a linear or mixed-integer programming formulation that can be solved globally and then validated on the real system.
Core Features & Use Cases
- LP/MILP formulation for discrete knobs: Builds decision-variable models (binary activation per option/slot) with mutual exclusivity and an explicit objective tied to your measured metric.
- Empirical constraint derivation: Converts observation-study results (quality floors, latency caps, failure-onset thresholds, rate-of-change triggers) into inequality constraints grounded in real runs.
- Linearization and validation loop: Supports common linearization patterns (max/min via auxiliary variables, bilinear relaxation, Big-M thresholding) and requires re-evaluating the LP-chosen configuration against the actual system, iterating until stable.
- Anti-correlation handling and fallbacks: Detects when constraints fight due to missing orthogonal dimensions and offers Pareto/tradeoff or grid-search fallback for small discrete spaces.
Quick Start
Use parm-tuning-as-lp to map your candidate discrete hyperparameter options into an LP/MILP with objective and constraints from recent evaluation data, solve it, and then validate the selected configuration on the real system before finalizing.
