Due to the complexity of metabolic regulation, first-principles models of bioreactor dynamics typically have built-in errors (structural and parametric uncertainty) which give rise to the need for obtaining relevant data through experimental design in modeling for optimization. A run-to-run optimization strategy which integrates imperfect models with Bayesian active learning is proposed. Parameter distributions in a probabilistic model of bioreactor performance are re-estimated using data from experiments designed for maximizing information and performance. The proposed Bayesian decision-theoretic approach resorts to probabilistic tendency models that explicitly characterize their levels of confidence. Bootstrapping of parameter distributions is used to represent parametric uncertainty as histograms. The Bajpai & Reuss bioreactor model for penicillin production validated with industrial data is used as a representative case study. Run-to-run convergence to an improved policy is fast despite significant modeling errors as long as data are used to revise iteratively posterior distributions of the most influencing model parameters. (C) 2012 Elsevier Ltd. All rights reserved.