Optimizer Card — bayesian-optimization v1.0.0

artifact_kind: optimizer-card artifact_name: bayesian-optimization version: 1.0.0 owner: Nous · agreement-search team last_reviewed: 2026-04-25 next_review: 2026-07-25 optimizer_kind: bayesian related_phase_179_tasks: [179.4.2.5]#

1. Identity#

• Optimizer name: bayesian-optimization
• Algorithm family: Bayesian optimization with Gaussian-process surrogate (RBF kernel) and Expected-Improvement / UCB / PI acquisition.
• Implementation source: libs/nous/agreement-search/src/bayesian-optimization.ts. No external GP library; Cholesky decomposition with automatic jitter is in-house.
• Intended use: expensive-to-score landscapes — frontier-model utility scoring, legal-reviewer sign-off, domain-simulator evaluations — where each fitness call costs cents to seconds and the GA / NSGA-II / MCTS budget of hundreds of evaluations is unaffordable.

2. Scoring rule#

• Utility aggregation: scalar fitness, caller-supplied.
• Acquisition functions: Expected Improvement, Upper Confidence Bound (β-tunable), Probability of Improvement.
• Fairness metrics reported: post-acquisition, via fairness-metrics.ts.
• Handling of hard constraints: every proposed point passes through filterCandidates before evaluation.
• Handling of uncertainty: explicit — the GP posterior IS the uncertainty model. Acquisition trades exploration vs exploitation using the GP's posterior variance.

3. Inputs#

• Variable space: typed parameter space via BoVariableSpec (int, continuous, boolean, enum); normalized to a unit hypercube [0,1]^d for GP modeling, decoded back when applying to the candidate.
• Initial design: scrambled Latin-hypercube-like sequence of initialDesignSize points before acquisition starts.
• GP kernel: isotropic RBF (squared-exponential) with caller-tuned rbfLengthscale and rbfAmplitude; Gaussian observation noise with variance observationNoiseVariance.
• Acquisition optimization: acquisitionSamples random samples → top-K refined by seeded Gaussian-perturbation local search.
• Time / iteration budget: total iterations + acquisition-iteration budget per step.
• Randomness: seeded mulberry32.

4. Outputs#

• Accepted candidate contract: the best evaluated candidate (argmax fitness).
• Diagnostics: GP posterior at every iteration, acquisition value history, per-iteration fitness, observation list.
• Uncertainty propagation: the final GP posterior is exposed for workbench rendering of "what we're still uncertain about".

5. Evaluation#

6. Known limitations#

• GP fit is O(n^3) in observations; practical budget ~150 observations per run before refit cost dominates. Sparse-GP variants are a roadmap item.
• RBF kernel assumes smooth fitness landscapes; discontinuous landscapes (hard threshold rewards) degrade performance.
• High-dimensional spaces (d > ~20) suffer from the curse of dimensionality; for these, prefer CP-SAT or shell to dimensionality- aware BO methods.

7. Guardrails#

• Candidate safety gate: filterCandidates before every fitness call; a filter failure is logged and the BO loop drops the point.
• Clause static validation: inside filterCandidates.
• Redline separation: hard constraints encoded into the variable space or via filterCandidates; the optimizer cannot trade them.
• Determinism guarantee: seeded; reproducible.

8. References#

• libs/nous/agreement-search/src/bayesian-optimization.ts
• Snoek, Larochelle, Adams (2012), "Practical Bayesian optimization of machine learning algorithms".
• Srinivas et al. (2010), "Gaussian Process Optimization in the Bandit Setting" (GP-UCB regret bound).
• docs/research/bibliography.md §bayesian-optimization