Box constraints for optimizers

[vincentqb]

🚀 Feature

Offer the option to set box constraints for optimizers.

Motivation

In some applications of reinforcement learning, falling outside of a region correspond to inadmissible actions.

Alternatives

One can omit the constraints, and hope that the optimal parameters found happen to be within the constraints, without any guarantees.

Additional context

This was suggested in #938 for torch.optim.lbfgs.

CC @bamos

[vincentqb]

Superset of #6564

[eric-tramel]

Thanks for this superset of the box-constraint feature request on LBFGS (i.e. LBFGS-B) @vincentqb . The contributed #938 was closed for being stale and also not the correct solution. However, this feature would be very useful for some inverse image reconstruction problems (used in privacy analysis of DCNNs, for example).

Should one attempt to resurrect the old #938 and address your concerns (correct implementation, align with current state of the code, etc.) in a followup PR specific for LBFGS, or are you looking for contributions only to address the more meta-challenge of box-constraints across all optimizers?

[nikitaved]

If there are no other volunteers, I would love to work on this after I am done with functions of a matrix. I used a box-constrained Newton in my research, based on the method by Bertsekas, which I had to implement from scratch, so it should not take that much time...

[nsakhanenko]

Has there been any progress in adding bounds to LBFGS (LBFGS-B) in pytorch? It would be very useful.