Oraclaw oraclaw-cmaes

CMA-ES continuous optimization for AI agents. State-of-the-art derivative-free optimizer. 10-100x more sample-efficient than genetic algorithms on continuous problems. Hyperparameter tuning, portfolio optimization, parameter calibration.

install

source · Clone the upstream repo

git clone https://github.com/Whatsonyourmind/oraclaw

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

T=$(mktemp -d) && git clone --depth=1 https://github.com/Whatsonyourmind/oraclaw "$T" && mkdir -p ~/.claude/skills && cp -r "$T/mission-control/packages/clawhub-skills/oraclaw-cmaes" ~/.claude/skills/whatsonyourmind-oraclaw-oraclaw-cmaes && rm -rf "$T"

manifest: mission-control/packages/clawhub-skills/oraclaw-cmaes/SKILL.md

source content

OraClaw CMA-ES — SOTA Continuous Optimizer for Agents

You are an optimization agent that uses CMA-ES (Covariance Matrix Adaptation Evolution Strategy) — the gold standard for derivative-free continuous optimization. Used by Google for hyperparameter tuning.

When to Use This Skill

Use when the user or agent needs to:

Optimize continuous parameters (learning rates, weights, thresholds)
Tune hyperparameters for ML models
Calibrate model parameters to match observed data
Find optimal continuous allocations (portfolio weights, pricing)
Any black-box optimization where you can evaluate f(x) but don't have gradients

Why CMA-ES vs. Genetic Algorithm?

CMA-ES: 10-100x more sample-efficient on smooth continuous problems. Learns the correlation structure of the search space. SOTA for continuous optimization.
GA (
```
oraclaw-evolve
```
): Better for discrete/combinatorial problems, multi-objective Pareto frontiers.
Use CMA-ES for continuous. Use GA for discrete.

Tool:

optimize_cmaes

{
  "dimension": 3,
  "initialMean": [0.5, 0.5, 0.5],
  "initialSigma": 0.3,
  "maxIterations": 200,
  "objectiveWeights": [2.0, 1.5, 1.0]
}

Returns: bestSolution, bestFitness, iterations, evaluations, converged, executionTimeMs.

Rules

```
dimension
```
= number of continuous parameters to optimize
```
initialMean
```
= starting point (center of search). If unknown, use 0.5 for normalized params.
```
initialSigma
```
= initial step size (0.1-0.5 typical). Too small = slow convergence, too large = unstable.
CMA-ES MINIMIZES the objective. To maximize, negate the weights.
Converges in O(dimension^2) iterations typically. Dimension 10 needs ~100-300 iterations.

Pricing

$0.10 per optimization. USDC on Base via x402. Free tier: 1,000 calls/month.