We propose a worldview to facilitate the discussion of how Monte Carlo simulation experiments are conducted in a digital twin context. We use the term Monte Carlo digital twin to refer to a digital twin in which a Monte Carlo simulation model mimics the physical twin. The worldview we propose encompasses Nelson and Schmeiser's classical worldview for simulation experiments and adds key components of a Monte Carlo digital twin including a predictor, a decision maker, and bidirectional interaction between the Monte Carlo digital twin and the physical twin. We delineate the proposed worldview components in three examples.
The performance of a CPU-only implementation of the restarted GMRES algorithm with direct randomized-SVD -based preconditioning has been analyzed. The method has been tested on a set of sparse and dense matrices exhibiting varying spectral properties and compared to the ILU(0) -based preconditioning. This comparison aims to assess the advantages and drawbacks of both approaches. The trade-off between iteration-to-solution and time-to-solution metrics is discussed, demonstrating that the proposed method achieves an improved convergence rate in terms of iterations. Additionally, the method's competitiveness with respect to both metrics is discussed within the context of several relevant scenarios, particularly those where GMRES-based simulation techniques are applicable.
We consider a short-term demand forecasting problem for semiconductor supply chains. In addition to observed demand quantities, order entry information is available. We compute a combinational forecast based on an exponential smoothing technique, a long short-term memory network, and the order entry information. The weights for the different forecast sources and parameters for exponential smoothing are computed using a genetic algorithm. Computational experiments based on a rich data set from a semiconductor manufacturer are conducted. The results demonstrate that the best forecast performance is obtained if all the different forecasts are combined.
This paper investigates the sample efficient exploration policy for asynchronous Q-learning from the perspective of uncertainty quantification. Although algorithms like $\varepsilon$-greedy can balance exploration and exploitation, their performances heavily depend on hyperparameter selection, and a systematic approach to designing exploration policies remains an open question. Inspired by contextual Ranking and Selection problems, we focus on optimizing the probability of correctly selecting optimal actions (PCS) rather than merely estimating Q-values accurately. We establish a novel central limit theorem for asynchronous Q-iterations, enabling the development of two strategies: (1) an optimization-based policy that seeks an optimal computing budget allocation and (2) a parameter-based policy that selects from a parametrized family of policies. Specifically, we propose minimizing an asymptotic proxy of Q-value uncertainty with regularization. Experimental results on benchmark problems, including River Swim and Machine Replacement, demonstrate that the proposed policies can effectively identify sample-efficient exploration strategies.
The advancement of Artificial Intelligence (AI) has accelerated the transformation of simulation from a tool for analysis and design into a dynamic partner for decision-making and operation. As AI systems become more capable of learning, reasoning, and adapting, simulation is evolving into an intelligent, autonomous, and predictive framework for exploring complex futures. This paper brings together future-oriented perspectives from simulation scientists and AI experts to discuss the current AI-simulation integration and both near-term and long-term outlooks for innovation, collaboration, and disruption.
