Re: reverse interval operations
Marco Nehmeier wrote:
P1788,
please find attached a motion on reverse interval operations as
defined in the Vienna proposal.
Any comments welcome
Marco
Marco, P1788,
Thank you for such a clear explanation of reverse mode.
I see the value it provides by helping to narrow unwanted interval pessimism
due to the extra constraint. For example, this can be useful in
branch-and-prune methods like interval Newton, or perhaps constraint
propagation.
In Eqs. (6) and (7), reverse mode does not appear to be a _new_ operation,
but rather a fusing of a basic arithmetic operation with an intersection
operation (the constraint).
From a practical perspective, what is the benefit of a custom circuit in
hardware dedicated to performing this fused operation? It seems a simple
interval processor with pipelining capabilities will perform the equivalent
non-fused operations very quickly.
In software, couldn't a user easily define these reverse operations as a
macro or inline function, using only the basic operations from Motion 6 and
an intersection operator?
Sincerely,
Nate