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.