Design space exploration of stochastic System-of-Systems simulations using adaptive sequential experiments

The complexities of our surrounding environments are becoming increasingly diverse, more integrated, and continuously more difficult to predict and characterize. These modeling complexities are ever more prevalent in System-of-System (SoS) simulations where simulation run times can surpass real-time and are often dictated by stochastic processes and non-continuous emergent behaviors. As the number of connections continue to increase in modeling environments and the number of external noise variables continue to multiply, these SoS simulations can no longer be explored with traditional means without significantly wasted computational resources.;This research will discuss the defining features of an SoS and many of the issues plaguing the SoS industry. Then, it will move to a literature review of the concepts currently used to explore design spaces, and finally, it will explore a set of two cascading research areas which will culminate in an adaptive sequential design of experiments for SoS simulations.;The first research area will investigate the key features to SoS and the attributes of these SoS which are important to be identified while exploring their simulations. To complete this investigation, first SoS properties are deduced from SoS's relationship to its super-class, complex systems. Second, following this examination, properties are further induced by investigating notional SoS simulations. From these two research avenues it will be discovered these spaces are nonparametric, conditionally variant, non-normally and non-identically distributed. Further, attributes of the output metrics are identified that will increase the likelihood of locating interesting regions of SoS simulations.;The knowledge and information gained from this first research focus is used in developing and comparing existing techniques capable of capturing SoS attributes. Several methods from the literature are compared on numerous stochastic mathematical problems and a single notional SoS simulation to determine their relative performance. From this comparison it will be shown that there are currently no methods capable of learning both the mean and variance of these complex spaces. Although the best method will be shown to be the MARS algorithm for generic high dimensional stochastic problems, it will be shown to be inadequate for SoS simulations.;Finally, these two research areas will enable the synthesis of an adaptive sequential algorithm capable of exploring stochastic simulations with emphasis on the attributes common to SoS. This final research area will determine strategically where to place points in the design space to improve its predictive capability. The final algorithm will be tested on an identical set of stochastic mathematical problems and the notional SoS simulation from the second research area, but will also include a published high dimensional SoS simulation. The final method will be shown to improve the exploration of stochastic simulations over existing methods by increased global accuracy, the number of simulations required to learn the space, and the computational speed.