Modelling of process systems with genetic programming
Thesis (MScEng (Process Engineering))--University of Stellenbosch, 2006.
Genetic programming (GP) is a methodology that imitates genetic algorithms, which uses mutation and replication to produce algorithms or model structures based on Darwinian survival-of-the-fittest principles. Despite its obvious po-tential in process systems engineering, GP does not appear to have gained large-scale acceptance in process engineering applications. In this thesis, therefore, the following hypothesis was considered: Genetic programming offers a competitive approach towards the automatic generation of process models from data. This was done by comparing three different GP algorithms to classification and regression trees (CART) as benchmark. Although these models could be assessed on the basis of several different criteria, the assessment was limited to the predictive power and interpretability of the models. The reason for using CART as a benchmark, was that it is well-established as a nonlinear approach to modelling, and more importantly, it can generate interpretable models in the form of IF-THEN rules. Six case studies were considered. Two of these were based on simulated data (a regression and a classification problem), while the other four were based on real-world data obtained from the process industries (three classification problems and one regression problem). In the two simulated case studies, the CART models outperformed the GP models both in terms of predictive power and interpretability. In the four real word case studies, two of the GP algorithms and CART performed equally in terms of predictive power. Mixed results were obtained as far as the interpretability of the models was concerned. The CART models always produced sets of IF-THEN rules that were in principle easy to interpret. However, when many of these rules are needed to represent the system (large trees), the tree models lose their interpretability – as was indeed the case in the majority of the case studies considered. Nonetheless, the CART models produced more interpretable structures in almost all the case studies. The exception was a case study related to the classification of hot rolled steel plates (which could have surface defects or not). In this case, the one of the GP models produced a singularly simple model, with the same predictive power as that of the classification tree. Although GP models and their construction were generally more complex than classification/regression models and did not appear to afford any particular advantages in predictive power over the classification/regression trees, they could therefore provide more concise, interpretable models than CART. For this reason, the hypothesis of the thesis should arguably be accepted, especially if a high premium is placed on the development of interpretable models.