Release notes

The format is based on Keep a Changelog.

v0.5.0 (unreleased)

Version 0.5 is a rewrite with a new interface. See Migrating from version 0.4 for the correspondence between the old and the new names.

Added

• PiecewiseLinear function objects, which are built independently of a model, with the function given first so that a do block can be used, and add the function to the model of their inputs when called, e.g. f(x). They replace piecewiselinear.
• Univariate functions given in pieces, e.g. with jumps between them, with PiecewiseLinear(f::Vector{<:Function}, x::Vector{<:AbstractVector}).
• Functions with several outputs, with three or more inputs, and with discontinuities, built from their segments with PWLFunction.
• Piecewise linear relaxations, i.e. bands between a lower and an upper bound, with UnivariatePWLRelaxation, formulated by IndependentBranching and DisaggregatedLogarithmic.
• The direction keyword with Epigraph and Hypograph, which only bound the output from one side.
• The methods IndependentBranching (Lyu, Hicks and Huchette), ConvexPieces, GreedyTriangleSelection and MultipleChoice for functions of any dimension.
• add_formulation returns a PWLFormulation with the variables and constraints it added, which can be deleted from the model again.
• Evaluation of functions with pwl(x...), and bivariate functions from a matrix of values.
• An extension interface with SOS2Method, TriangleSelectionMethod and formulate_pwl!, see Extending the package.
• Documentation with a manual, tutorials and an API reference.

Changed

• Methods are values of their own types instead of symbols, e.g. Logarithmic() instead of :Logarithmic.
• The default triangulation of bivariate functions is K1Triangulation, and the default method depends on the triangulation: K1, UnionJack or GreedyTriangleSelection.
• The default method of univariate functions is MultipleChoice for discontinuous functions and DisaggregatedLogarithmic for relaxations, instead of Logarithmic, which supports neither.
• The triangle selections of SixStencil and NineStencil leave out levels that separate nothing, so they use fewer binary variables on small grids and on triangulations with a single diagonal direction.
• OptimalTriangleSelection and OptimalIndependentBranching take a time limit, cache their results, and fall back to the greedy selection.
• The inputs of a function may be affine expressions or constants.

Removed

• The Celaya formulations. Use ZigZagInteger for a general integer formulation.