jayantmukho
commented
4 years ago Thanks for all the questions on the shape optimization. It helps us make improvements to the code to make it easier to use.
You are correct in that the adjoint gradients determine the direction that the optimizer moves the design variables. But this direction can (and usually does) change during the optimization process as it explores different parts of the design space. So the design variables won't always move in the same direction.
Sometimes the optimizer won't take the best step in the direction of descent. In this case, the objective function can increase. The optimizer will then try again in the same direction but with a different (usually smaller) step size. If it does achieve a reduction in the objective function, it will usually run the adjoints again to get an updated gradient direction. The objective + adjoint iterations are called major iterations and are what is output by the SLSQP algorithm. This process is more complicated when there are constraints involved.
So in summary, no it is not necessary (or likely) that the objective function will always decrease through an optimization process
auzbaig
pcarruscag
Although I am minimizing my objective function, its value appears to rise in some designs while running a shape optimization. I look at my objective function value from the history_project.dat file.
It makes sense for it to move in only one direction because that's what the adjoint gradients will ask it to do, however, I don't know if this rule is actually always satisfied in shape optimization.