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
Find which configuration works best for the base implementation and using which hyperparameters. For example, should 'base' sampling or 'sampling_region' be implemented by default?
Then, set automatic hyperparameters based on configuration. These should be overwritten when the user provides any specific input unless there's an explicit clash. For example, good initial guesses would be:
Set initial radius to cover 10-20% or 50% of the search space for local and global flavour
Set beta_red to accuracy**(1/N) or accuracy**(2/N), where accuracy would be for example 0.001 or 0.01, could be related to min_radius for TIS or TIP routines
Set N_min_s to be for example a tenth of the budget, depending on local or global flavour or depending on PLS routine
These automatic hyperparameters should be specific to the 'flavour' or configuration that the user desires. Some ideas for configurations:
'Exploration': Use 'base' sampling in combination with some exploration routine, potentially 'exploit_explore' or 'sampling_region'
'Multistart': User should be able to choose between TIS and TIP, defaulting to either
'Expensive_local': Implies that the user wants to get to some local optimum as quickly as possible. Use local method, and small trust regions
'Expensive_global': Not sure if this makes sense, but potentially a routine with a lot of exploration, potentially more risky
'Safety': Implement feasible sampling
'structure' or 'high-dimensional': Use CUATRO_PLS or CUATRO_low when implemented, think of heuristics for n_e or n_pls
'ill-behaved': should be similar to exploration or multistart
'well-behaved' or 'convex': can use larger trust regions since the approximation should be quite good. not sure if the local or global method would work better
The text was updated successfully, but these errors were encountered:
Find which configuration works best for the base implementation and using which hyperparameters. For example, should 'base' sampling or 'sampling_region' be implemented by default?
Then, set automatic hyperparameters based on configuration. These should be overwritten when the user provides any specific input unless there's an explicit clash. For example, good initial guesses would be:
These automatic hyperparameters should be specific to the 'flavour' or configuration that the user desires. Some ideas for configurations:
The text was updated successfully, but these errors were encountered: