NonconvexBayesian

NonconvexBayesian provides Gaussian-process surrogate-assisted (Bayesian) optimization for the Nonconvex ecosystem. Expensive objective and constraint functions (flagged with :expensive) are replaced by GP surrogates whose acquisition model is solved by any Nonconvex optimizer to propose the next query point; the true functions are then evaluated at the candidate to update the surrogates.

This documentation is a fragment of the Nonconvex documentation and is also built standalone.

Features

• BayesOptAlg meta-algorithm: wraps any Nonconvex optimizer (BayesOptAlg(IpoptAlg()), BayesOptAlg(NLoptAlg(...)), ...).
• Acquisition functions: LCB (default), GPUCB (with a β(t) schedule), EI and PI; constraint surrogates always use LCB for conservative feasibility.
• Zero-order (ZeroOrderGPSurrogate), first-order (FirstOrderGPSurrogate) and co-kriging (CoKrigingSurrogate, joint value+gradient) surrogates.
• Direct-Sobol or acquisition-based initialization, acquisition multistart with duplicate-point filtering, and TuRBO-style local trust-region restarts for high-dimensional problems.
• Expensive constraints replaced by conservative (LCB) surrogates, with equality constraints transformed into 2-block inequalities (see Constraints).
• Full protocol integration: GenericResult-style results with fcalls, convergence state, callbacks, verbose logging, keep_all query trajectories and reset! support for meta-algorithm nesting.
• Solver capability declarations forwarded to the sub-solver (solver_capabilities).

See Quick start for a complete example.