Crate rlx_bbo 

Black-box optimization + FMQ/QGBS search (domain-agnostic f64 objectives).

For compiled flow-map policies and RLX-graph FMQ training, use rlx-rl.

Re-exports

pub use bo::Acquisition;

pub use bo::BoConfig;

pub use bo::bo;

pub use gp::GpPosterior;

pub use gp::Kernel;

pub use gp::cholesky;

Modules

acquisition

Acquisition functions for Bayesian optimisation.

bo

Top-level Bayesian optimisation loop.

gp

Gaussian-process posterior with RBF and Matérn kernels.

sampling

Low-discrepancy / space-filling samplers for design-of-experiments.

tpe

Tree-structured Parzen Estimator (Bergstra et al. 2011, Optuna).

Structs

AdamOptConfig

AdamOptResult

BboSolution

Bbox

CmaesConfig

EsConfig

GraphOptConfig

GraphOptResult

GraphOptSpec

Parameter bundle for adam_opt_graph.

LinearFlowMap

One-step flow map X_{0,1}(a₀) = a₀ + W·a₀ + b (linear MVP analogue).

LinearSurrogate

Affine Q(x) ≈ b + w·x trained by ridge regression on logged trajectories.

PsoConfig

QSteerConfig

QgbsConfig

TrajectoryRecord

One design point along an optimization or placement run.

Enums

GraphOptError

Constants

DEFAULT_KAPPA

Stability constant for normalized gradient steps (paper κ₁).

Functions

adam_opt_1d

adam_opt_graph

Compile grad_with_loss on fwd and run Adam.

adam_opt_nd

append_jsonl

cmaes

Diagonal CMA-ES. Minimizes f.

diagonal_flow_pairs

Extract (action, target_velocity) pairs for diagonal flow matching: v* ≈ a₁ − a₀.

eta_eff_twin

Adaptive trust radius from twin critic disagreement (paper eq. 13, batch-normalized δ).

find_param_node

Resolve a single Op::Param node by name.

find_param_nodes

Resolve param nodes in the same order as names.

finite_diff_grad

Centered finite-difference gradient of f at x.

fit_from_trajectory_jsonl

fit_linear_surrogate

Ridge fit: minimize ‖Xw − y‖² + λ‖w‖² with bias column.

fmq_surrogate_step

FMQ-style online step using surrogate grad: x ← x − η ∇Q / ‖∇Q‖.

gradcheck_graph

Central-difference gradcheck for fwd w.r.t. optimize params.

load_flow_map

load_jsonl

load_surrogate

one_plus_one_es

pso

q_guided_beam_search

Q-guided beam search (QGBS): renoise → optional trust steer → keep top beam by f.

q_steered_search

Q-steered search: anchor at x_ref, trust-region gradient steps + random exploration.

q_steered_search_twin

Q-steered search with twin objectives: cheap (e.g. prescreen) and expensive (e.g. ngspice).

q_steered_search_with_grad

Q-steered search with a caller-supplied gradient (e.g. rlx AD); skips finite differences.

random_search

save_flow_map

save_surrogate

search_by_method

Run random search, Q-steer, or QGBS from a string tag ("bbo", "qsteer", "qgbs").

train_from_jsonl

Offline train from JSONL trajectories; returns flow map + training MSE.

trust_region_q_step

One trust-region step toward lower f (minimization): x_ref − η · ∇f / (‖∇f‖ + κ).