%0 Report %1 hinchcliffe:1996:c2GPcpsm %A Hinchliffe, Mark %A Willis, Mark %A Hiden, Hugo %A Tham, Ming %C UK %D 1996 %K algorithms, genetic programming %T A comparison of two Genetic Programming Algorithms Applied to Chemical Process Systems Modelling %X Previous work by McKay et al (1996a,b,c) has shown that the Genetic programming (GP) methodology can be successfully applied to the development of non-linear steady state models of industrial chemical processes. Although a GP algorithm can identify the relevant input variables and evolve parsimonious system representations, the resulting model structures tend to contain little or no information relating to the mechanisms of the process itself. In this respect, the performance of the GP methodology is comparable to that of other black-box modelling techniques such as neural networks. Chemical process systems are often extremely complex and non-linear in nature. Phenomenological models are time consuming to develop and can be subject to inaccuracies caused by any simplifying assumptions made. Consequently, mechanistic models are costly to construct; an aspect which would make an automated procedure highly desirable. Phenomenological models are usually derived by applying the laws of conservation of mass, energy and momentum to the system. An examination of a number of steady-state mechanistic models shows that they tend to be made up of distinct sub-groups which, when added together, give the overall model structure. In the search for an automatic model generating algorithm, it would be extremely useful if the GP methodology could be used to identify these sub-groups. This could potentially enhance the GP algorithm's ability to evolve accurate chemical process models and also help to reveal hidden process knowledge. To achieve this goal, the standard GP algorithm used by McKay et al (1996a) was modified to accommodate the multiple gene model structure. The multiple gene structure was introduced by Altenberg (1994) in an attempt to enhance the learning capabilities of GA and GP algorithms. The work was inspired by the observation that, in nature, genetic information is stored on more than one gene. To demonstrate the feasibility of this new approach, real world examples are used to compare the performance of the algorithm with that of the standard GP algorithm.

@techreport{hinchcliffe:1996:c2GPcpsm, abstract = {Previous work by McKay et al (1996a,b,c) has shown that the Genetic programming (GP) methodology can be successfully applied to the development of non-linear steady state models of industrial chemical processes. Although a GP algorithm can identify the relevant input variables and evolve parsimonious system representations, the resulting model structures tend to contain little or no information relating to the mechanisms of the process itself. In this respect, the performance of the GP methodology is comparable to that of other black-box modelling techniques such as neural networks. Chemical process systems are often extremely complex and non-linear in nature. Phenomenological models are time consuming to develop and can be subject to inaccuracies caused by any simplifying assumptions made. Consequently, mechanistic models are costly to construct; an aspect which would make an automated procedure highly desirable. Phenomenological models are usually derived by applying the laws of conservation of mass, energy and momentum to the system. An examination of a number of steady-state mechanistic models shows that they tend to be made up of distinct sub-groups which, when added together, give the overall model structure. In the search for an automatic model generating algorithm, it would be extremely useful if the GP methodology could be used to identify these sub-groups. This could potentially enhance the GP algorithm's ability to evolve accurate chemical process models and also help to reveal hidden process knowledge. To achieve this goal, the standard GP algorithm used by McKay et al (1996a) was modified to accommodate the multiple gene model structure. The multiple gene structure was introduced by Altenberg (1994) in an attempt to enhance the learning capabilities of GA and GP algorithms. The work was inspired by the observation that, in nature, genetic information is stored on more than one gene. To demonstrate the feasibility of this new approach, real world examples are used to compare the performance of the algorithm with that of the standard GP algorithm.}, added-at = {2008-06-19T17:35:00.000+0200}, address = {UK}, author = {Hinchliffe, Mark and Willis, Mark and Hiden, Hugo and Tham, Ming}, biburl = {https://www.bibsonomy.org/bibtex/2a4f4fa4a1da480faad101c9dd843ad28/brazovayeye}, broken = {http://lorien.ncl.ac.uk/sorg/paper10a.ps}, institution = {Chemical Engineering, Newcastle University}, interhash = {1fb53cdbe93553ec4b0417bbaf681a06}, intrahash = {a4f4fa4a1da480faad101c9dd843ad28}, keywords = {algorithms, genetic programming}, note = {Extended Abstract, submitted to: ICANNGA '97, Norwick, UK}, notes = {MSword postscript not camptible with unix}, size = {7 pages}, timestamp = {2008-06-19T17:41:27.000+0200}, title = {A comparison of two Genetic Programming Algorithms Applied to Chemical Process Systems Modelling}, year = 1996 }