Result

optimize returns a BayesOptResult with the GenericResult-style fields:

• minimizer, minimum: the best feasible solution found,
• iter: the number of BO iterations run (one expensive evaluation each),
• maxiter_reached: whether the loop stopped at maxiter without converging,
• tol: the tolerances (ftol, ctol) used,
• convstate: the convergence state of the last candidate (feasibility, Δx, Δf, relΔf, converged),
• fcalls: the number of expensive-function evaluations performed during the call (the initial point is not counted),
• n_restarts: the number of TuRBO-style trust-region restarts,
• initial_x, optimizer,
• sub_result: the result of the last surrogate (acquisition) solve,
• surrogates: the trained surrogates (their X/y fields hold the query history),
• trajectory: the per-iteration (x, f, feasible, iter) records when keep_all = true.

A callback passed to optimize is invoked once per BO iteration with a Solution carrying the candidate, the objective value, the constraint violations and the convergence state. verbose = true logs one line per iteration.

Workspace

The BayesOptWorkspace underlying an optimization keeps the original model, the surrogate sub-workspace, the surrogates, the callback and the query history. It is created by the standard NonconvexCore Workspace protocol and supports reset!:

w = Workspace(m, BayesOptAlg(IpoptAlg()), x0, options = options)
NonconvexCore.reset!(w)   # rebuild the surrogates from scratch

This makes the workspace reusable and nestable in other meta-algorithms (such as multistart drivers).