Traditionally, modelling tasks involve the building of mathematical equations which can best
describe the underlying process. Such a modelling practice normally requires a deep
understanding of the systems under investigation, hence the reason why it is often referred to
as knowledge-driven modelling. On the contrary, knowledge extraction from data (or datadriven
modelling), inspired principally from artificial intelligence techniques, is based on
limited knowledge of the modelling process and relies on the data describing the input and
output mappings. Such a process is able to make abstractions and generalisations of the
process and plays often a complementary role to knowledge-driven modelling.
The Fuzzy Rule-Based System (FRBS) has been found more appealing for such a knowledge
extraction process, compared to other ‘black-box’ modelling techniques, due to its ability of
providing human understandable knowledge. However, such interpretability is only semiinherent
in the FRBS. Without a special caution one can easily end up with a FRBS with
equally good predictions as those given by the ‘black-box’ modelling methods, while on the
other hand with equally bad interpretability. Hence, extracting a transparent (interpretative)
FRBS is reckoned to be of a multi-objective nature with often conflicting outcomes, which
gives the rationale of using bio-inspired optimisation paradigms, more specifically, Artificial
Immune Systems, in this research project. In a bid to further improve the overall predictive
performance, especially for the scatter and uncertain data set, an error correction scheme is
proposed so that one can compensate the original predictive model via the predicted error.
The proposed immune optimisation framework was tested extensively using several
benchmark problems and was compared with other salient techniques. Consistent better
performances were obtained. The immune based modelling approach was tested using a set of
benchmark problems, and was further applied to different real data sets, viz. Tensile Strength
(TS), Elongation and Reduction of Area (ROA), taken from the steel industry, which are all
featured by high dimensional, nonlinear and sparse data spaces. Results show that the
ii
proposed modelling approach is capable of eliciting not only accurate but also transparent
FRBSs. Such a transparent FRBS establishes the required predictions of the mechanical
properties of materials, which on the one hand can help metallurgists to further understand
the underlying mechanisms of alloys processing, and on the other hand will automate and
simplify their design. Charpy toughness (impact energy) as a special data set featured by
scatters and uncertainties was used to validate the proposed error correction mechanism and
proved its validity.
The project is part of the research activities which are currently conducted in the Institute for
Microstructural and Mechanical Process Engineering: The University of Sheffield
(IMMPETUS).
