One of the most important issues of empirical model based batch process optimal control is that the calculated optimal control profile can degrade very significantly when applied to the actual process because of model-plant mismatches. "Optimal on the model" can be quite different from "optimal on the process". To address this issue, this paper presents a reliable optimal control method where the optimization objective function includes an additional term to penalize wide model prediction confidence bounds at the end point of a batch. Bootstrap aggregated neural networks are used to model a batch polymerization reactor from limited batches of process operational data. The model can predict the number-average molecular weight, weight-average molecular weight, and monomer conversion at several points during a batch from the batch recipe and reactor temperature profile. A further advantage of bootstrap aggregated neural network models is that model prediction confidence bounds can be obtained. By penalization of wide model prediction confidence bounds at the end point of a batch, the calculated optimal control profile is much more reliable in the sense that, when it is applied to the actual process, the degradation in the control performance is limited.