> My main point was that mutation is not there to escape local optima, although I should have been clearer about the efficacy of crossover for that purpose.
Mutation is absolutely there for the primary purpose of escaping local optima. If mutation's job was only to climb the gradient, then you would just use a much more efficient gradient ascent method. By having random mutation, you are effectively saying you want to stay in the known-good region most of the time, but occasionally explore a new area even if it goes against the perceived gradient.
Mutation is absolutely there for the primary purpose of escaping local optima. If mutation's job was only to climb the gradient, then you would just use a much more efficient gradient ascent method. By having random mutation, you are effectively saying you want to stay in the known-good region most of the time, but occasionally explore a new area even if it goes against the perceived gradient.