What problem does it solve? Tuning hyperparameters or calibrating simulators often means optimizing a function you can only evaluate, with no usable gradients, and ad-hoc runs produce results nobody can reproduce or compare. This Skill runs covariance matrix adaptation evolution strategy (CMA-ES) optimization with explicit parameters, declared stop rules, and a full convergence trace so every result is evidence rather than anecdote. ## Core Features & Use Cases - Reproducible run records: Every run logs the objective characterization, initial mean and sigma, population size, seed, stop-rule set, and noise handling before evaluating. - Disciplined stop rules and restarts: Runs terminate only on declared conditions (evaluation budget, target fitness, stall limit, sigma floor, wall clock), and restarts with doubled population happen only when declared up front. - Fair benchmark comparisons: Compare optimizers on sphere, Rosenbrock, Rastrigin, or Ackley functions only under equal evaluation budgets and identical starting regions. - Use Case: A researcher tuning six hyperparameters of an expensive simulation declares a 2,000-evaluation budget and seed, runs CMA-ES, and receives the best point, the stop rule that fired, and a monotone-checked convergence trace a colleague can re-run to verify. ## Quick Start Ask the assistant to minimize your objective function with CMA-ES using a declared evaluation budget, seed, and bounds, and to return the convergence trace and stop rule that fired.
