rlx-bbo 0.2.6

Black-box optimization and FMQ/QGBS search (trust-region Q-guidance, arXiv:2605.12416).
Documentation

rlx-bbo

Generic black-box optimization and Flow Map Q-Guidance (FMQ) / QGBS for any f64 objective — no simulator bindings.

Graph-based offline/online FMQ training (compiled MLP actor + twin critic on rlx-ir) lives in rlx-rl.
Reference JAX code: /Users/Shared/q-guided-flow-map-policies.

Modules

Module Role
q_guidance Trust-region step, q_steered_search, q_guided_beam_search
twin Cheap vs expensive twin-η search
surrogate Linear ridge critic from trajectory JSONL
flow_map Diagonal linear flow-map offline training
trajectory JSONL logging for offline datasets
cmaes Separable CMA-ES

rlx-eda

Depend on rlx-bbo directly (workspace path ../rlx/rlx-bbo). EDA-specific prescreen↔ngspice twins and harness JSON helpers stay in eda-fmq.

License

GPL-3.0-only.