Skillsbench mpc-horizon-tuning

Selecting MPC prediction horizon and cost matrices for web handling.

install

source · Clone the upstream repo

git clone https://github.com/benchflow-ai/skillsbench

Claude Code · Install into ~/.claude/skills/

T=$(mktemp -d) && git clone --depth=1 https://github.com/benchflow-ai/skillsbench "$T" && mkdir -p ~/.claude/skills && cp -r "$T/tasks/r2r-mpc-control/environment/skills/mpc-horizon-tuning" ~/.claude/skills/benchflow-ai-skillsbench-mpc-horizon-tuning && rm -rf "$T"

manifest: tasks/r2r-mpc-control/environment/skills/mpc-horizon-tuning/SKILL.md

source content

MPC Tuning for Tension Control

Prediction Horizon Selection

Horizon N affects performance and computation:

Too short (N < 5): Poor disturbance rejection
Too long (N > 20): Excessive computation
Rule of thumb: N ≈ 2-3× settling time / dt

For R2R systems with dt=0.01s: N = 5-15 typical

Cost Matrix Design

State cost Q: Emphasize tension tracking

Q_tension = 100 / T_ref²  # High weight on tensions
Q_velocity = 0.1 / v_ref²  # Lower weight on velocities
Q = diag([Q_tension × 6, Q_velocity × 6])

Control cost R: Penalize actuator effort

R = 0.01-0.1 × eye(n_u)  # Smaller = more aggressive

Trade-offs

Higher Q	Effect
Faster tracking	More control effort
Lower steady-state error	More aggressive transients

Higher R	Effect
Smoother control	Slower response
Less actuator wear	Higher tracking error

Terminal Cost

Use LQR solution for terminal cost to guarantee stability:

P = solve_continuous_are(A, B, Q, R)