Steering Parameters¶
This part only relevant for constrained problems
steering_l1_model¶
Boolean value. Default value: True
Determines whether or not the one norm (the standard choice) or the infinity norm is used for the total violation measure in steering, which in turn affects the predicted violation reduction.
steering_ineq_margin¶
Non-negative real value. Default value: 1e-6
Sets the margin of feasibility for problems that only have inequality constraints. For such problems, steering can be adaptively enabled instead of applied at every iteration. When the inequality constraints are all at least steering_ineq_margin away from being active, steering will be disabled. Setting steering_ineq_margin to zero means that steering will only be applied when one or more inequality constraints are active ( >= 0) while setting steering_ineq_margin to inf means that steering will be applied on every iteration. NOTE: this parameter has no effect if equality constraints are present and steering will always be enabled every iteration.
steering_maxit¶
Positive integer. Default value: 10
If the default steering direction is not predicted to adequately promote progress towards feasibility, steering_maxit is the max number of times to iteratively lower the penalty parameter and recompute a hopefully better search direction which does adequately promote progress towards feasibility. Setting this to higher values may more aggressively lower the penalty parameter per NCVX iteration.
steering_c_viol¶
real value in (0,1). Default value: 0.1
Determines the fraction of predicted violation reduction for a search direction to be considered as adequately promoting progress towards feasibility. If a search direction’s predicted violation reduction does not meet this target, the penalty parameter will be lowered and a new search direction will be computed. Setting this to higher values may more aggressively lower the penalty parameter per NCVX iteration.
steering_c_mu¶
real value in (0,1). Default value: 0.9
When steering lowers the penalty parameter in order to compute a a search direction which promotes progress towards feasibility, it does by progessively lowering the penalty parameter by factors of steering_c_mu. For any given NCVX iteration, the penalty parameter can be shrunk up to a factor of steering_c_mu to the power of steering_maxit. Setting this to lower values may more aggressively lower the penalty parameter per NCVX iteration.