Parameter Constraints
Last updated
Last updated
Parameter Constraints can be added to an experiment definition to search for the optimum within a limited region of parameter space. Once parameter constraints are defined, all the configurations generated by SigOpt are guaranteed to satisfy them.
Linear inequality constraints can be used to define restrictions that depend on multiple parameters. Multiple linear constraints can be defined at the same time where each constraint must affect more than one parameter. The region where all the constraints are satisfied is called the feasible region. SigOpt will find the optimum inside the feasible region, while guaranteeing that no points are placed outside the feasible region. In the case of multiple constraints, SigOpt also checks the existence of the feasible region at experiment creation. For constraints that cannot be explicitly expressed as linear inequalities (e.g., nonlinear or blackbox), refer to the Metric Constraints documentation on defining constraints through thresholding on metric values.
A SigOpt Experiment with linear inequality constraints is defined by the expression:
The weight of each parameter as well as the threshold
inequality value can be set at experiment creation.
If the constraints are defined such that a feasible region does not exist, or where a constraint affects only one parameter (i.e. the constraint is encoded in the bounds of the respective parameter), SigOpt will return an API Error.
Nonzero weights are included only for the parameters involved in each constraint, e.g., observe that alpha
is not involved in the second constraint.
While the constraint types are less_than
or greater_than
, these are not strict inequalities, and we allow for equality as well. This will likely be more relevant to users defining constraints on int
parameters.
The experiment type must be offline
or random
The constraints can be defined in terms of any set of parameters of type double
or int
Mixed variable constraints are not allowed, i.e., a constraint involving both double
and int
parameters
Integer constraints can be defined on at most 10 unique int
parameters
For Experiments with conditional parameters, constraints can only be defined for unconditional parameters
The constraint definitions and constraint parameters cannot be updated during the experiment