morphogen.optimization.mma

Method of moving asymptotes (Svanberg 1987, 2002).

Each iteration replaces the problem

minimize f0(x) subject to fi(x) <= 0 (i = 1..m), xmin <= x <= xmax

by a convex separable approximation built on moving asymptotes low < x < upp and solves it by a primal-dual interior-point method, as in Svanberg’s notes “MMA and GCMMA, two methods for nonlinear optimization” (2007), with artificial variables y_i of cost c_i y_i + d_i y_i^2 / 2 that make every subproblem feasible.

class morphogen.optimization.mma.MMA(n: int, m: int, xmin, xmax, *, move: float = 0.5, asyinit: float = 0.5, asyincr: float = 1.2, asydecr: float = 0.7, c: float = 1000.0)[source]

Bases: object

Optimizer state: call update() with the current point, objective and constraint values and gradients; it returns the next point.

reset() → None[source]

Forget the asymptotes after a change of the problem (a continuation step): asymptotes tightened by past oscillations would hold the design in place.

update(x, f0, df0, fval, dfdx) → ndarray[source]