One possibility of getting around the issue of getting stuck in local minima via gradient descent is a technique called stochastic gradient descent, further explored here: http://proceedings.mlr.press/v80/kleinberg18a/kleinberg18a.pdf