- author = "Mark P. Hinchliffe",
- title = "Dynamic Modelling Using Genetic Programming",
- school = "School of Chemical Engineering and Advanced Materials, University of Newcastle upon Tyne",
- year = "2001",
- address = "UK",
- month = sep,
- keywords = "genetic algorithms, genetic programming, MOGA, MOGP, SOGP",
-
URL = "
http://www.cs.ucl.ac.uk/staff/W.Langdon/ftp/papers/hinchliffe:Thesis.pdf",
- broken = "http://www.ncl.ac.uk/ceam/postgrad/pg-theses.htm",
-
URL = "
http://ethos.bl.uk/OrderDetails.do?did=5&uin=uk.bl.ethos.391407",
- size = "205 pages",
-
abstract = "Genetic programming (GP) is an evolutionary algorithm
that attempts to evolve solutions to a problem by using
concepts taken from the naturally occurring
evolutionary process. This thesis introduces the
concepts of GP model development by applying the
technique to steady-state model evolution. A variation
of the algorithm known as the multiple basis function
GP (MBF-GP) algorithm is described and its performance
compared with the standard algorithm. Results show that
the MBF-GP algorithm requires significantly less
computational effort to evolve models of comparable
accuracy to the standard algorithm. The steady-state
algorithm is then modified to enable the evolution of
dynamic process models. Three case studies are used to
demonstrate algorithm performance and show how the
MBF-GP algorithm produces performance benefits similar
to those observed in the steady-state modelling work. A
comparison with neural networks reveals that GP is able
to match the accuracy of the network predictions but is
more expensive computationally. However, a significant
advantage of the GP algorithm is that it can
automatically evolve the time history of model terms
required to account for process characteristics such as
the system time delay.
The model development process is not simply a case of reducing the error between the predicted and actual process output. The parallel nature of GP means that it is ideally suited to solving multi-objective problems. The MBF-GP algorithm is modified to incorporate a Pareto based ranking scheme that allows models to be compared using multiple performance criteria. The ranking scheme allows preference information in the form of goals and priorities to be specified in order to guide the search towards the desired region of the search space. Two case studies are used to demonstrate the performance of this technique. The first example uses the multi-objective algorithm to improve the parsimony of the evolved model structures. The second example demonstrates how a set residual correlation tests can be combined and used as an additional performance measure. In each case, the multi-objective algorithm performs significantly better than the single objective version. In addition, the inclusion of preference information overcomes some of the difficulties associated with conventional Pareto ranking and produces a greater number of acceptable solutions.",
-
notes = "{"}the results do not provide sufficient evidence to
suggest that GP will become as widely used as neural
network modelling techniques.{"} page
160.
uk.bl.ethos.391407",
