You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
AEPsych models currently require lower bounds and upper bounds of parameters to be passed into their constructors, but the model itself should not have to worry about the bounds of the space; that's the responsibility of the strategy. All the model really needs to know is the dimensionality of the space. Removing the lb and ub from the constructor would reduce a lot of boilerplate. Wherever a model method requires self.lb or self.ub, we should be able to refactor it so that the method takes those as parameters that can be passed in by the strategy.
The text was updated successfully, but these errors were encountered:
ub/lb/dim arguments in the [common] block are all being deprecated.
ub/lb will no longer be defined in the [common] block, instead they will be defined with upper_bound/lower_bound in parameter specific blocks.
dim should never be defined in config files anymore and should be inferred based on all of the parameters. This also implies parnames must always be set now.
Important places to check their usage:
Accessing ub/lb to build objects from the config currently rely on ub/lb in [common] as fallback, this needs to be changed to build ub/lb tensors from parameter-specific settings
Using dim from the config is used to setup benchmarking test cases, this should be done with benchmarking specific behavior instead. A good indicator of this are in test configs that do not have parnames set.
AEPsych models currently require lower bounds and upper bounds of parameters to be passed into their constructors, but the model itself should not have to worry about the bounds of the space; that's the responsibility of the strategy. All the model really needs to know is the dimensionality of the space. Removing the lb and ub from the constructor would reduce a lot of boilerplate. Wherever a model method requires self.lb or self.ub, we should be able to refactor it so that the method takes those as parameters that can be passed in by the strategy.
The text was updated successfully, but these errors were encountered: