Investigates online control of fermentation with Arthrobacter viscosus. These bacteria secrete penicillin acylase, a key enzyme in pharmaceutical industry. The growth of more cells during fermentation results in more enzyme. Enzyme activity and cell growth are sensitive to pH, so pH control during batch fermentation is very important. Two peristaltic pumps, supplying acidic and basic solutions, are used, and a 4-4-1 recurrent backpropagation neural net (RBPN) is used for adaptive control because of its long-term identification ability. The transfer function x/(1+|x|) is used. The deviation of the pH from the set point of pH 7 is the input node of the network controller. Its output node is the predicted flow rate of the pump for next control time interval. The model was operated by two phases. During the first, it was set as the process model and trained by a fixed set of online acquired data. During the second, it acted as a predictor, the predicted control action was hence obtained. To enhance effective computation of this network, the number of training data was limited. A moving window of size 15 for supplying training data was determined for each learning and applied for the online learning. Good results were obtained, including a maximum optical density of 6.7 at the end of the fermentation.