Expand description
Optimisation: the objective and optimiser interfaces, the optimiser registry, and the algorithms.
Every optimiser minimises an Objective through the same Optimizer trait and reports to the same
ProgressSink, which can also cancel the run. All of them stop when the target fidelity is reached, return the
best point they evaluated, and never see threads: parallelism lives inside the objective.
A new algorithm is one type implementing Optimizer plus one entry in optimizer and optimizer_names;
Monitor gives it progress reports, cancellation, the target check and best-point tracking.
Structs§
- Bobyqa
- BOBYQA, Powell’s bound-constrained quadratic-model trust-region method. Initial trust radius 0.5.
- Cobyla
- COBYLA, Powell’s linear-approximation trust-region method. Initial trust radius 1.0, as in scipy.
- Console
Sink - A sink that prints an iteration table to standard output, every
every-th row. - Eval
- One evaluation of the objective.
- Iteration
Report - One row of progress.
- LBfgs
- Limited-memory BFGS with a Moré–Thuente line search.
- Monitor
- Bookkeeping every optimiser shares: progress reports, the stop checks and the best point.
- Newton
Cg - Line-search Newton-CG with exact Hessian-vector products.
- Newton
Exact - Newton’s method in a trust region on the exact Hessian, with the subproblem solved by Steihaug’s truncated conjugate gradients, so indefinite Hessians are handled.
- NoProgress
- A sink that ignores everything.
- Optim
Result - What an optimisation produced.
- Recording
Sink - A sink that keeps every report.
- RunControl
- Run settings shared by every optimiser.
Enums§
- Derivatives
- What an optimiser asks the objective for.
- Exit
- Why an optimisation ended.
Traits§
- Objective
- A function to minimise.
- Optimizer
- An optimisation algorithm.
- Progress
Sink - Something that watches an optimisation and can stop it.
Functions§
- optimizer
- The optimiser with configuration name
name. - optimizer_
names - Every optimiser name, in the order the interface lists them.