Is there a subset of these constrained optimisation techniques that work when the constraint bounds are unknown before evaluation time? This applies to certain physical engineering problems where, due to the complexity of the environmental system you're modelling, you don't necessarily know what combination of input parameters lead to a non-viable solution. That is, a given solution might be numerically valid but the objective function could determine it to be non-viable due to not satisfying engineering constraints (e.g. it vibrates too fast).

Just adding a penalty to the objective works. Another alternative is to consider a two-dimensional lexicographic objective (total constraint violation, actual objective), so that reducing constraint violations is always preferable, and the actual objective is only compared between solutions that are equally good in terms of constraint violations.

Annnnnnnd you've added gradients to the gradient free optimizer. Every shovel is a hammer if try hard enough.