Abstract base classes

Optimistix is fully extendable. It provides a number of abstract base classes ("ABCs") which define the interfaces for custom solvers, custom line searches, etc.

Solvers

• Custom minimisers may be created by subclassing optimistix.AbstractMinimiser.
  • If your minimiser is derived from Optax, then you can simply use optimistix.OptaxMinimiser.
• Custom least squares solvers may be created by subclassing optimistix.AbstractLeastSquaresSolver.
• Custom root finders may be created by subclassing optimistix.AbstractRootFinder.
• Custom fixed-point solvers may be created by subclassing optimistix.AbstractFixedPointSolver.

Info

For creating least squares solvers like optimistix.GaussNewton or optimistix.LevenbergMarquardt, then optimistix.AbstractGaussNewton may be useful.

For creating gradient-based minimisers like optimistix.BFGS and optimistix.GradientDescent, then optimistix.AbstractGradientDescent may be useful.

In each case, they offer a general way to combine a search and a descent.

Line searches, trust regions, learning rates etc.

These may be defined by subclassing optimistix.AbstractSearch. See also the introduction to searches and descent

Descent directions (steepest descent, Newton steps, Levenberg--Marquardt damped steps etc.)

These may be defined by subclassing optimistix.AbstractDescent. See also the introduction to searches and descent

Adjoints

These denote custom autodifferentiation strategies. These may be defined by subclassing optimistix.AbstractAdjoint.

Norms

Any function PyTree -> non-negative real scalar may be used as a norm. See also the norms page.

Nonlinear CG variants

Any function (Y, Y, Y) -> scalar may be used to define a variant of nonlinear CG. See optimistix.polak_ribiere as an example.