Figure 3
From: Agents for sequential learning using multiple-fidelity data
Overview of how uncertainty is used to balance the use of low and high fidelity data in the implementation of Algorithm 2 of Section S2. When a data point x is called for as an experimental candidate, the above flowchart describes the decision-making process for which data source to use. After a data point x is selected for measurement, two conditions are checked: (1) if the corresponding (e.g. with the same formula) low-fidelity measurement has already been made or (2) if the uncertainty associated with the high fidelity measurement is low enough (below a threshold \(\beta\)). If either is true, the high-fidelity measurement is taken. If neither are true, then the agent must consider the trade-off between low-fidelity and high-fidelity data. The agent thus considers how a low fidelity data point would affect the current ordering of the predicted figure of merit associated with all candidates. If it would alter the ranking by more than \(\gamma\), the low-fidelity measurement is taken. If not, the high-fidelity measurement is taken. Note that \(\beta\) and \(\gamma\) are user-defined hyperparameters explained in detail in Section S2.
