Induction of fuzzy rules from chemical process data using Growing Neural Gas and Reactive Tabu search methods

[摘要] ENGLISH ABSTRACT: The artificial intelligence community has developed a large body of algorithms that can beemployed as powerful data analysis tools. However, such tools are not readily used inpetrochemical plant operational decision support. This is primarily because the models generatedby such tools are either too inaccurate or too difficult to understand if of acceptable accuracy.The Combinatorial Rule Assembler (CORA) algorithm is proposed to address these problems.The algorithm uses membership functions made by the Growing Neural Gas (GNG) radial basisfunction network training technique to assemble internally disjunctive, Oth -order Sugeno fuzzyrules using the nongreedy Reactive Tabu Search (RTS) combinatorial search method.An evaluation of the influence of CORA training parameters revealed the following. First, forcertain problems CORA models have an attribute space overlap that is one third of their GNG-generatedcounterparts. Second, the use of more fuzzy rules generally leads to better modelaccuracy. Third, decreased swap (or move) thresholds do not consistently lead to more accurateand / or simpler models. Fourth, utilisation of moves rather than swaps during rule antecedentassembly leads to better rule simplification. Fifth, consequent magnitude penalisation generallyimproves accuracy, especially if many rules are built. Variance of results is also usually reduced.Sixth, employing Yu rather than Zadeh operators leads to improved accuracy. Seventh, use of theGNG adjacency matrix significantly reduces the combinatorial complexity of rule construction.Eighth, AlC and BIC criteria used find the right-sized model exhibited local optima. Last, theCORA algorithm struggles to model problems that have a low exemplar to attribute ratio.On a chaotic time series problem the CORA algorithm builds models that are significantly (withat least 95% confidence) more accurate than those generated using multiple linear regression(MLR), CART regression trees and multivariate adaptive regression splines (MARS). However,only the RTS component models are significantly more accurate than those of the GNG and k-means(RBF) radial basis function network methods. In terms of complexity, the CORA modelswere significantly simpler than the CART and RBF models but more complex than the MLR,MARS and multilayer perceptron models that were evaluated. Taking all results for this probleminto account, it is the author's opinion that the drop in accuracy (at worst 0.42%) of the CORAmodels, because of membership function merging and rule reduction, is justified by the increasein model simplicity (at least 22%). In addition, these results show that relatively intelligibleif. .. then ...fuzzy rule models can be built from chemical process data that are competitive (interms of accuracy) with other, less intelligible, model types (e.g. multivariate spline models).

[发布机构] Stellenbosch University