Unfold Partial-Least Squares (u-PLS) is a modeling method successfully applied to batch-process monitoring and end quality prediction. This method is integrated in a self-tuning optimization algorithm, based on extremum-seeking control. The optimization is driven by the gradient obtained by means of an adaptive u-PLS model. Since this is an empirical model, no first-principles based knowledge of the process is necessary. Heuristic rules are used to constrain the gradient taking into account nonlinearity and unknown causes of variability. Extensions to model the variability in initial conditions, to optimize several performance indices, and to handle inequality constraints are presented. The optimization algorithm is tested on a complex comprehensive simulated-process model for the fed-batch cultivation of Sacharomyces cerevisiae. Results show the performance and versatility of the proposed approach, as well as its robustness to process changes. (c) 2007 American Institute of Chemical Engineers.