OptionalceilingThe greatest penalty the steering may reach: a growth that would pass it is not made, and the round goes on with the step at the penalty reached, as it does when STEERING.increases are spent. Unbounded when absent; read only with steer.
OptionalfromThe penalty to start at, a solve continued from another's result (IResult.penalty); the first when absent.
The penalties, ascending, each tried while the constraints are not met at the one before.
OptionalradiusThe region's first half-width, a solve following another of the same kind (IResult.radius); ISchedule.trust when absent.
Convexifications per penalty at most.
The region's schedule.
OptionalstationaryThe least reduction of the violation the caller can tell, in the
violation's units: its own feasibility tolerance. Where the steering's
step of the violation alone lies inside its region (not at its edge),
it is the linearised violation's own least; reducing the violation by
no more than this while the iterate violates by more, it shows the
iterate a first-order stationary point of the violation, an infeasible
one that no later round leaves, and the solve stops there with reason
infeasible. Absent, a solve runs its rounds; read only with steer.
OptionalsteerWhether the penalty is steered (Byrd, Nocedal and Waltz 2008, "Steering
exact penalty methods for nonlinear programming", Optim Methods Softw
23:197; Algorithm 5.1, p. 11 of the 2007 technical report): each round,
where the step at the penalty leaves the linearised constraints
violated (IStep.met), the step of the violation alone (the model
without the objective, their (5.10)) sets the progress a step can make;
the penalty grows by STEERING.increase until the step meets the
linearised constraints where that one does, or else reduces their
violation by at least STEERING.feasibility of what that one does
(their (5.13), p. 10), and then until the model's decrease is at least
STEERING.merit times the penalty times the step's reduction of the
violation (their (5.14), p. 10). The penalty is then tied to the
problem at every iterate, so a start the constraints' linearisation
cannot satisfy is not pulled off them by the objective, and the steps
keep their progress on feasibility. With it, penalties gives only the
first penalty and from a continued one. False when absent.
Minimise from an iterate. What a solve is run with.