This is a minor PR where I've updated a broken link to Nocedal's optimization book.
After reading through literature on filter methods, I have one small comment on the side (which I can raise as a separate issue if appropriate):
I have a concern regarding whether requested results of this RFR would have much scientific merit, as it appears that Filter Methods find their main use for solving the constraint optimization formulation and not a multi-objective optimization setting as described in this RFR (that is, the MORL setting: where the task is attempting to optimize a collection of reward functions).
Recall that, filter methods essentially reduces the constraint optimization problem to a bi-objective optimization task (where one objective is the solution quality and another is a quantification of how much said solution violates the given constraints). We can run any iterative optimization process and then track the Pareto front for this bi-objective optimization task (that is, actively considering whether each intermediate iterate of said process belongs on the Pareto front or not), and ultimately favouring solutions that do well in the constraint objective once the optimization procedure terminates (since non-feasible solution are not viable). If we took this idea plainly and brought it to a general MO setup: it is evident that tracking Pareto dominant solutions of a MO process does not add much scientific value in of itself.
I think the only real benefit from this RFR is seeing how well the creative "bells and whistles" that were independently evolved within the filter methods community translates within a MORL setting.
This is a minor PR where I've updated a broken link to Nocedal's optimization book.
After reading through literature on filter methods, I have one small comment on the side (which I can raise as a separate issue if appropriate):