Cost-Aware Active Learning for Feasible Region Identification

Design space exploration for engineering design involves identifying feasible designs that satisfy design specifications, often represented by feasibility constraints. To determine whether a design is feasible, an expensive simulation is required. Therefore, it is crucial to find and model the feasible region with as few simulations as possible. Model-based Active learning (AL) is a data-efficient, iterative sampling framework that can be used for design space exploration to identify feasible regions with the least amount of budget spent. A common choice for the budget is the number of (sampling) iterations. This is a good choice when every simulation has an equal cost. However, simulation cost can vary depending on the design parameters and is often unknown. Thus, some regions in the design space are cheaper to evaluate than others. In this work, we investigate if incorporating the unknown cost in the AL strategy leads to better sampling and, eventually, faster identification of the feasible region.