There are things like NAS (neural architectural search) but all you are doing is just growing the search space and making the optimization problem much harder. Typically you do the architectural optimization by hand, using heuristics and past experiments as guidance.