Paramus — Active-Learning Loop

Overview

Given completed-experiment results, fit/update a surrogate, propose the next most informative experiments, and decide continue-vs-stop. The value is the iterate-until-converged strategy, not one fit.

Guidance vs execution: surrogate fitting and acquisition come from tested Paramus tools. Do not hand-roll a Gaussian process or acquisition function in ad-hoc code.

When to Use

• Sequential optimization where each round chooses the next best experiments
• Expensive experiments where sample efficiency matters Do not use for: fixed full-factorial designs (use the campaign planner) or pure post-hoc analysis.

Workflow

Prior results + objective + candidate space + budget
  ↓ 1. Update surrogate model      → fit to all results so far
  ↓ 2. Assess uncertainty          → where is the model unsure?
  ↓ 3. Acquisition → next batch     → EI / UCB / max-info, honoring constraints
  ↓ 4. Stop check                  → converged? budget spent? target met?
  ↓ 5. Emit next experiments        → batch + rationale + provenance

Procedure

1. Discover tools via search/get_schema ("Bayesian optimization", "active learning", "surrogate").
2. Update the surrogate with all results to date; report fit quality.
3. Quantify predictive uncertainty across the candidate space.
4. Apply the acquisition strategy to propose the next batch within constraints.
5. Evaluate stop criteria (convergence, budget, target); recommend continue or stop.
6. Emit the next experiments with rationale and provenance.

Domain Judgment

• Choose acquisition to match intent: exploration (UCB) vs exploitation (EI) vs pure information gain.
• Respect batch constraints and feasibility — proposals must be physically runnable.
• Report surrogate fit quality each round; a poor model makes proposals untrustworthy.

Fallbacks

• Optimization tool unavailable → deliver current surrogate summary + a manual next-step suggestion, clearly scoped.
• Endpoint unreachable → stop and report; do not fabricate acquisition scores.