irishtier

Technically combining stochasticity with gradient descent is not pure gradient descent.

As for the domain knowledge - this is not specific to gradient descent either, genetic algorithms and their problem encoding can be adapted to a domain to minimise the search space considered - therefore improving performance of the algorithm.

Totally obsurd. Gradient descent will 'almost' always get stuck in a local optima. The point of evolutionary algorithms, or any naturally inspired algorithm for that matter, is that it does not tend to converge upon local optima and has the possibility of moving out of this space towards the global optima. The fact that most problems it is applied to are NP-Complete means that it is unlikely it will find the optimal solution in polynomial time.