Reliable Decison Making in Digital Twins

National Science Foundation, AFOSR

Motivation & Problem Description

  • Digital twins continuously exchange information with physical systems. Unlike conventional simulation models, they are updated with streaming data and send predictions and decisions back to the physical system.
  • This closed-loop interaction creates a central methodological challenge: decisions must be made and revised while the underlying model is still being learned and updated.
  • The project addresses uncertainty from limited physical data, changing model estimates, decision-dependent observations, and constrained simulation and data-collection resources.

Objective

  • Develop a unified decision-making perspective for digital twins that connects simulation optimization, stochastic optimal control, reinforcement learning, and risk-aware decision making.
  • Determine how a digital twin should combine information across model updates, allocate limited data and simulation resources, and update decisions as new evidence arrives.

Solution

  • Model the digital twin as an adaptive data–model–decision loop: physical observations update the twin, the updated model predicts consequences, and a decision module selects actions that affect future states and observations.
  • Develop and synthesize methods for data-driven simulation optimization under input uncertainty, streaming data, contextual information, and joint resource allocation.
  • Connect digital twins with stochastic control and reinforcement learning, including Bayesian approaches that update models, policies, and risk assessments as information accumulates.
  • Use distributionally robust and Bayesian risk formulations to manage model uncertainty and the sim-to-real gap.

Impact

  • Provides a common framework for designing digital twins that adapt decisions while learning, rather than treating the twin as a fixed simulation model.
  • Clarifies when additional physical data, additional simulation, model updates, or risk protection are most valuable for reliable decisions.
  • Supports future digital twins that are adaptive, computationally efficient, risk-aware, and capable of combining automation with human oversight.
Reliable Decison Making in Digital Twins
click image to enlarge
Dynamic decision making with a digital twin: observations update the model, the model supports a decision, and implementation
changes the physical system and future data.
References:
  • Zhou, E. (2024). Data-Driven Simulation Optimization in the Age of Digital Twins: Challenges and Developments. Proceedings of the 2024 Winter Simulation Conference, 31–45.
  • Wu, D., Wang, Y., and Zhou, E. (2024). Data-Driven Ranking and Selection under Input Uncertainty. Operations Research, 72(2), 781–795.
  • Wang, Y., and Zhou, E. (2023). Bayesian Risk-Averse Q-Learning with Streaming Observations. NeurIPS.

ISyE Faculty PI

Enlu Zhou

Enlu Zhou

Fouts Family Professor

ISyE ProfilePersonal Website