| Summary: | The two traditional approaches to modelling can be characterised as the development of mechanistic models from 'first principles' and the fitting of statistical models to data. The so-called 'hybrid approach' combines both elements within a single overall model and is thus composed of a set of mass balance constraints and a set of kinetic functions. This thesis considers methodologies for building hybrid models of bioprocesses. Two methodologies were developed, evaluated and demonstrated on a range of systems of simulated and experimental systems. A method for inferring models from data using support vector machines was developed and demonstrated on 3 experimental systems a Murine hybridoma shake flask cell culture, a Saccharopolyspora erythraea shake flask cultivation and a 42L Streptomyces clavuligerus batch cultivation. On the latter system the method produced models of similar accuracy to previously published hybrid modelling work. While support vector machines have been widely applied in other contexts this method is novel in the sense that there are no previously published papers on the use of support vector machines for kinetic modeling of bioprocesses. On 50 randomly created dynamical systems it was shown that the hybrid models produced using the support vector machine methodology were generally more accurate than those developed using feed forward neural networks and that could not be distinguished from models produced using a more computationally demanding method based round genetic programming. Additionally a Bayesian framework for hybrid modelling was developed and demonstrated on simple simulated systems. The Bayesian approach requires no interpolation of data, can cope with missing initial conditions and provides a principled framework for incorporating a priori beliefs. These features are likely to be useful in practical situations where high quality experimental data is difficult to produce.
|
|---|
