orion
orion copied to clipboard
Handle broken trials in Hyperband, ASHA and EvolutionaryES
Following #602, broken trials we never be set a lie objective (although it does not make much difference as HB, ASHA & EES where meant to be used with StubParallelStrategy(stub_value=None)
). The algorithms should handle the broken trials themselves.
We could rely on ParallelStrategy module to set a bad objective, but I think they should instead consider the broken trial as an invalid one that does not count in the rung count or population count (as done for PBT in #705). This way an unlucky bracket with many broken trials will not be starved but instead sample additional trials to compensate.