Big data analytics are used to explore vast quantities of transactional data, such as customer purchase histories, medical records, or cell phone usage, to discover anticipated and unanticipated relationships that are useful for prediction and understanding. Dynamic stochastic simulations are used by engineers and management scientists to design and improve industrial, service and financial systems that are subject to risk. Simulations also generate large quantities of synthetic transactional data, but current practice is to first distill these data into high-level summary performance measures, and then to discard it, making it difficult to evaluate risk or predict actual system behavior. The success of data analytics in business and industry will lead simulation users to expect the same sort of fine-grained analysis from their simulations, and if they cannot obtain it they may conclude that simulation is irrelevant. This Grant Opportunity for Academic Liaison with Industry (GOALI) research project will provide a foundation and proof-of-concept first steps toward a data analytics treatment of dynamic, stochastic simulation, by considering simulation as data analytics for systems that do not yet exist. The result will be better and more robust system design and control decisions for business and industry. Collaboration with industrial partner SAS Institute will insure that the research is broadly relevant and the results are implemented.
Three core topics are planned: simulation analytics as a precursor to system control, simulation analytics for comprehensive comparisons of system designs, and simulation analytics via dynamic metamodels. The focus is on a post-simulation analysis that is facilitated by retaining the simulated sample paths, sample paths that may have been generated by a simulation experiment designed to achieve a specific narrow objective such as system optimization. New visualization and statistical analysis methods will be created to allow simulation users to solve the kinds of problems that data scientists routinely address, but in the simulation context where the data are dynamic, time-dependent sample paths, rather than customer instances. This will require new statistical methods for both supervised and unsupervised learning. Problems that do not typically arise in field data analytics, like the comparison of alternative system designs that are exposed to the same source of simulated uncertainty, will also be addressed.
