Microbial fermentations involving two or more kinds of competing cells and operating under realistic conditions are difficult to monitor, model and optimize by model-based methods. They deviate from ideal behavior in two significant aspects: incomplete dispersion in the broth and the influx of disturbances. The appraoch here has been to optimize the filtered noise and dispersion on-line through neural networks. This method has been applied to the fed-batch production of streptokinase (SK). The culture has two kinds of cells - active (or productive) and inactive - and their growth is inhibited by the substance and the primary metabolite (lactic acid). Using simulated data, the fermentation was optimized by a system of three neural networks, updated continually during successive time intervals. Such sequential optimization with dynamic filtering of inflow noise generated better cell growth and SK activity than static optimization and evan an ideal fermentation.