Information relaxation in stochastic optimal control