Learning-Based Portfolio Policies as Approximations to Risk-Sensitive Stochastic Control