Neural and Neuro-Fuzzy Approaches to Support ``Intelligent'' Scientific Problem Solving

Scientific computing uses computers, especially Higll Performance Computing (HPC) systems, to solve complex mathematical equations which model pllysica! p!lcnomena. Using these systems now requires expert knowledge in a variety of fields of computer science, such as parallel computing and numerical methods. This often makes application scientists, who lla,'c tllC domain cxpcclisc to devise the mathematical models, unable to use the power of HPC systems. Tbe object of problem solving environments (PSEs) is to create software systems that bide the details and complexity of the system [rolll the users, and to allow them to deal with a high level, abstract entity that understands the application domain "Ianguagen. This requires approximate reasoning tcchniques to automate much of nl1merical and parallel computing, as well as to interpret the users input. Over the past several years, we llavc developed PYTllIA, all "intelligent" computational assistant to achieve this goal. III tltis paper, wc describe the COllllcdionist tcdLliiqucs used in developing PYTHIA. Specifically, we dwcuss backpropagation based systems, as well as hybrid neura-fuzzy systems whicll we have developed and used. We also compare tile performance of these alternative approaches wHll each other, as well as with naive classifiers.