Grey Wolf Optimisation is the search half of the loop, the part that proposes designs. It keeps a population of candidate frames and, each iteration, ranks them by score and names the best three the alpha, beta and delta. The rest of the pack then updates toward those three leaders. The design space is far too large to check by hand, so instead of enumerating it the search circles inward on the good regions.
The one dial that matters is the balance between exploration and exploitation: range widely enough to escape a merely-decent design, but tighten enough to actually converge. Tuning that balance is what a published paper came out of. Early in a run the pack ranges; later it closes on the leaders. Get the schedule wrong and the search either wanders forever or freezes on the first local optimum it meets.
The search never judges its own proposals. It hands each candidate to the solver and takes back a score, which keeps the roles clean: the search explores, the solver decides.