Neural approximations for optimal control of nonlinear stochastic systems