文摘
Biological processes are non-deterministic systems. In general, models dealing with microbial pathways and microbial physiology are exceedingly complex, and as it is almost impossible to measure intracellular concentrations on-line, they normally have too many uncertain parameters that are difficult to evaluate. It is important, therefore, to generate modeles based on on-line measured variables, as these can be used in process control and on-line optimization of biological systems. Neural networks, which offer a data-driven modeling approach, are well suited for the above purpose, having good generalization and prediction capabilities. Neural networks include, among others, the feedforward-backpropagation and recurrent types. Localized networks such as Radial Basis Function (RBF) networks have found use in on-line process control due to the fact that they require less computational time. Another type of network, Cascade Correlation (CC), has been developed to generate the optimal structure for a particular application. Important factors to be considered in selecting and using neural networks in biotechnology are: proper scaling of data, selection of an appropriate network structure¡ªincluding suitable choices of input and output variables, and the purpose of the network. In this paper, the above issues and different networks are studied in context with different microbiological systems. Case studies dealing with fuel alcohol production using renewable biomass from agricultural wastes by fermentation with Zymomonas mobilis and recombinant Escherichia coli, and preliminary results for the production of monoclonal antibody using hybridoma cells, are examined. The results of these studies indicate that RBF networks are unsuitable when extrapolation is desired. Among Multi-layer Perceptron (MLP) networks, Recurrent Neural networks and Cascade Correlation networks provide the best prediction capabilities.
