@inproceedings{mercure:2001:AIChE, title = {Empirical Emulators for First Principle Models}, address = {Reno Hilton}, author = {Peter Kip Mercure and Guido F. Smits and Arthur Kordon}, booktitle = {AIChE Fall Annual Meeting}, month = {6 November}, url = {http://www.aiche.org/conferences/techprogram/paperdetail.asp?PaperID=2373&DSN=annual01}, year = {2001}, biburl = {http://www.bibsonomy.org/bibtex/23e91785362a357e1204178965f384500/brazovayeye}, abstract = {Empirical emulators mimic the performance of first principle models by using various data-driven modeling techniques. The driving force for developing empirical emulators is the push for reducing the time and cost for new product development. Empirical emulators are especially effective when hard real-time optimization of a variety of complex fundamental models is needed. The increased robustness of the modern data-driven techniques (analytic neural networks, support vector machines, genetic programming, etc.) is a reliable basis for accurate representation of fundamental models and gives many opportunities for effective synergy between these two key modeling approaches. The main schemes for building empirical emulators are discussed in the paper. Several contemporary techniques for robust empirical emulator design are explored including analytic neural networks, recurrent neural networks and Genetic Programming (GP), and the capabilities of the proposed approach are illustrated with a case study for a simple first principle model. A key feature of empirical emulators is that the training data for empirical model building is generated by design of experiments from first principle models called simulators. This allows a high degree of freedom for development of reliable data-driven models. The most obvious scheme for implementation of empirical emulators is as accelerator of computational time for fundamental models (the gain is 103 to 105 times faster). Another possible scheme is to use the empirical emulator as an estimator of fundamental model performance. Of special importance to on-line optimization is a scheme using the empirical emulator to integrate different types of fundamental models (steady-state, dynamic, fluid, kinetic, thermal, etc). Most of the known empirical emulators are implemented as {"}classical{"} neural networks based on back-propagation learning algorithm. Their property of being universal approximators is a key theoretical result for successful emulation. At the same time {"}classical{"} neural networks suffer from a number of problems like: long computational time for training, convergence to local minima, sensitivity to weight generalization, too many tunable parameters, etc. These problems put serious limitations on the quality of the developed empirical model, increase development time, and require experienced model developers. An alternative empirical emulator based on analytic neural networks is described in the paper. A key advantage of analytic neural networks is that the function to be optimized is a quadratic function of the weights of the hidden-to-output layer error and has one global optimum. It is no longer possible to get stuck in local minima and the learning algorithm is not iterative. As a result, the data-driven modeling process is significantly reduced and the developed empirical models are parsimonious. Of special importance to empirical emulator's performance is the ability of analytic neural networks to deliver multiple-model solution with confidence limits. Empirical emulators with confidence limits are aware of their own performance which is essential for any data-driven model application, especially in real-time. In the case of emulating process dynamics a different type of recurrent neural networks are needed. Recurrent networks are neural networks with one or more local or global feedback loops. The application of feedback enables neural networks to acquire state representations, making them suitable for emulation of dynamic fundamental models. A proper structure of an empirical emulator to mimic dynamic behavior is based on a recurrent version of the analytic neural networks.}, organisation = {AIChe}, notes = {American Institute of Chemical Engineers, 3 Park Ave, New York, N.Y., 10016-5991, U.S.A.}, keywords = {algorithms, genetic programming } } @inproceedings{kotancheck:2002:gecco, title = {Evolutionary Computing in {Dow Chemical}}, address = {New York, New York}, author = {Mark Kotanchek and Arthur Kordon and Guido Smits and Flor Castillo and R. Pell and M. B. Seasholtz and L. Chiang and P. Margl and P. K. Mercure and A. Kalos}, booktitle = {GECCO-2002 Presentations in the Evolutionary Computation in Industry Track}, editor = {Lawrence ``Dave'' Davis and Rajkumar Roy}, month = {11-13 July}, pages = {101--110}, year = {2002}, biburl = {http://www.bibsonomy.org/bibtex/2fb2e0aa9145e2450a958922e12d9bb61/brazovayeye}, organisation = {ISGEC}, notes = {powerpoint slides? Diverse subsets from chemical libraries. Soft sensors in intelligent alarm processing. Polypropylene structure-property relationships. non-linear DOE (design of experiments) using GP/GENPRO See also \cite{kordon:2002:gecco}}, keywords = {PSO, SVM algorithms, genetic machines, networks, neural particle programming, support swarm, vector } } @inproceedings{pmw05social, title = {The Social Semantics of LiveJournal FOAF: Structure and Change from 2004 to 2005}, author = {John C. Paolillo and Sarah Mercure and Elijah Wright}, booktitle = {Proceedings of the 1st Workshop on Semantic Network Analysis at the ISWC 2005 Conference}, day = {7.}, editor = {G. Stumme and B. Hoser and C. Schmitz and H. Alani}, month = {November}, pages = {69 -- 80}, year = {2005}, biburl = {http://www.bibsonomy.org/bibtex/20c71fdb7e05d9a1e45ad66061721cf05/stumme}, keywords = {FCA OntologyHandbook blog blogging seminar2006 } } @inproceedings{pmw05social, title = {The Social Semantics of LiveJournal FOAF: Structure and Change from 2004 to 2005}, address = {Galway, Ireland}, author = {John C. Paolillo and Sarah Mercure and Elijah Wright}, booktitle = {Proceedings of the 1st Workshop on Semantic Network Analysis at the ISWC 2005 Conference}, editor = {Gerd Stumme and Bettina Hoser and Christoph Schmitz and Harith Alani}, month = {November}, pages = {69 -- 80}, year = {2005}, biburl = {http://www.bibsonomy.org/bibtex/2bdfacf00e7525f28499bec855c5a495e/jaeschke}, day = {7.}, keywords = {blog blogging seminar2006 } } @article{journals/nar/LafontaineMP97, title = {Update of the viroid and viroid-like sequence database: addition of a hepatitis delta virus RNA section.}, author = {Daniel A. Lafontaine and Stéphane Mercure and Jean-Pierre Perreault}, journal = {Nucleic Acids Research}, number = {1}, pages = {123-125}, url = {http://dblp.uni-trier.de/db/journals/nar/nar25.html#LafontaineMP97}, volume = {25}, year = {1997}, biburl = {http://www.bibsonomy.org/bibtex/24250cee0d807d2f45e19c99001919baf/dblp}, description = {dblp}, date = {2005-06-09}, keywords = {dblp } } @article{journals/nar/LafontaineMPP98, title = {The viroid and viroid-like RNA database.}, author = {Daniel A. Lafontaine and Stéphane Mercure and Véronique Poisson and Jean-Pierre Perreault}, journal = {Nucleic Acids Research}, number = {1}, pages = {190-191}, url = {http://dblp.uni-trier.de/db/journals/nar/nar26.html#LafontaineMPP98}, volume = {26}, year = {1998}, biburl = {http://www.bibsonomy.org/bibtex/25601e3250c7dc4da1bb787312dd262ce/dblp}, description = {dblp}, date = {2003-11-25}, keywords = {dblp } }