Digital Twin-Based Optimization with Predictive Simulation Learning

Across domains and practices, computer simulations are a common tool that engineers, economists, decision makers, and others utilize to project real world impacts we might realize from controllable systems interacting with users and their environments. Simulations are preferable over other quantitative modelling methods when the relationship between systems, users, environments, and measurable outcomes is extremely complex, which almost always involves some aspect of randomness. While simulations give us the capability to quantitatively model these complex interactions, running these models requires a large amount of computational overhead. In fact, simulations may be precluded from use because this overhead is too large, where “too large” implies that the level of certainty in the observed model output required cannot be achieved in the time allotted for the given decision using the physical resources available. When the time constraints of the user cannot be accommodated by the simulation to execute enough replications to yield a solution that is at the precision required, the simulation cannot be effectively leveraged for decision making. The research presented in this dissertation seeks to find ways to reduce the computational overhead required by simulations to achieve a given level of output precision. A general class of algorithms referred to as simulation optimization algorithms addresses these issues, and the primary contribution of this research is the introduction of three novel simulation optimization algorithms; the Sequential Allocation using Machine-learning Predictions as Light-weight Estimates (SAMPLE) algorithm, the Robust Optimal Sampling (ROS) algorithm, and the Epsilon Optimal Sampling (EOS) algorithm. The algorithms introduce new ways to achieve better computational results when running simulations by integrating low-fidelity machine learning estimates with on-line simulation observations.

Author

• Goodwin, Travis

• Bayesian
• decision analysis
• OCBA
• optimal computing
• simulation
• simulation optimization

Identifier

hdl:1920/13999

OAI identifier oai:identifier

oai:MARS:1920/13999