Skip to content

Latest commit

 

History

History
77 lines (57 loc) · 3.17 KB

NEWS.md

File metadata and controls

77 lines (57 loc) · 3.17 KB
Raw

User visible changes in OptimPackNextGen package

Version 0.4.2

  • Fix loading of Zygote extension with Julia ≤ 1.8.

Version 0.4.1

  • In Powell's methods (COBYLA, NEWUOA, and BOBYQA):
    • Preserve scaling factors from being garbage collected while calling the C code.
    • Variables, bounds, and scaling factors can be specified in more flexible ways than dense vector of C-double values.

Version 0.4.0

  • New methods Brent.fmax and Brent.fmaxbrkt.

  • Methods for univariate functions Brent.fzero, Brent.fmin, Brent.fminbrkt, BraDi.minimize, BraDi.maximize Step.minimize, Step.maximize, and Step.search have changed as follows:

    • x and f(x) may have units.
    • The floating-point type for computations is not Float64 by default but determined from the types of the specified numerical arguments.
    • The number of function calls has been appended to the tuple returned by these methods.
    • The keyword period of BraDi.minimize and BraDi.maximize has been replaced by peridoc which is a Boolean, the period being given by the distance between the extreme values of x.
    • Brent.fzero yields the 5-tuple (x,fx,lo,hi,nf) with x the estimated solution, fx = f(x) the corresponding function value, lo and hi the lower and upper bounds for the solution, and nf the number of calls to f.
    • Step.minimize, Step.maximize, Step.search, BraDi.minimize, and BraDi.maximize return a 5-tuple (xm,fm,lo,hi,nf) with xm the position of the global optimum, fm = f(xm) the corresponding function value, lo and hi the lower and upper bounds for the exact solution, and nf the number of function calls. This is similar to fmin and fmax.
    • Step.minimize, Step.maximize, and Step.search no longer have a maxeval keyword. The number of function evaluations thus only depends on the requested accuracy for the solution.

  • The STEP method for finding a global minimum or maximum of an univariate function f(x) has been improved in many respects:

    • x and f(x) may have units.
    • The numerical precision used by the algorithm is automatically defined by the floating-point types of x and f(x).
    • Speed-up computations (by a factor of two) by simplifying the storage of trials.
    • Keywords have different names: atol and rtol specify absolute and relative tolerances for the precision of the solution, aboost and rboost specify absolute and relative boost parameters to define the function value to aim at.

Version 0.3.1

  • Fix missing methods to handle upper bounds.

Version 0.3.0

  • Automatic-differentiation now possible with Zygote.

  • Fix termination of SPG when the function value no longer change.

  • Keyword verb in vmlmb and spg can be an integer to print information every verb iteration.

  • Additional points can be specified in fmin. This replaces unexported methods fmin0, fmin1, fmin2, and fmin3 which have been suppressed.

Version 0.1

  • Use LazyAlgebra package for vectorized operations and linear conjugate gradient.