Sequential Information-Theoretic and Reification-Based Approach for Querying Multi-Information Sources

摘要

While the growing number of computational models available to designers can solve a lot of problems, it complicates the process of properly using the information provided by each simulator. It may seem intuitive to select the model with the highest accuracy, or fidelity, as decision makers want the greatest degree of certainty to increase their efficacy. However, high-fidelity models often come at a high computational expense. While comparatively lacking in veracity, low-fidelity models do contain some degree of useful information that can be obtained at a low cost. We propose a sequential method to use this information to generate a fused model with superior predictive capability than any of its constituent models. Our methodology estimates the correlation between each model using a model reification approach that eliminates the observational data requirement. The correlation is then used in an updating procedure whereby uncertain outputs from multiple models may be fused together to better estimate some quantity or quantities of interest. These ingredients are used in a decision-theoretic manner to query from multiple information sources sequentially to achieve the maximum knowledge about the fused model in as few information source evaluations as possible with minimum cost.