Nonlinear model predictive control (NMPC) is an effective method for optimal operation of batch processes. Most dynamic models however contain significant uncertainties. It is therefore important to take these uncertainties into account in the formulation of the open-loop MPC problem to prevent infeasibilities or worse performance. An issue of such formulations is the disregard of feedback in the predictions, which leads to overly conservative control actions. The introduction of feedback through parametrized control policies is one way to solve this issue. In this work we compare the performance of affine feedback policies against more complex policies given by radial basis function networks. We incorporate these feedback policies into a polynomial chaos based stochastic NMPC algorithm to gauge their efficiency. The parameters of the feedback policies are either determined online by the NMPC algorithm or are pre-computed offline.