Reinforcement Learning-Based Sequential Batch-Sampling for Bayesian Optimal Experimental Design

Abstract Engineering problems that are modeled using sophisticated mathematical methods or characterized by expensive-to-conduct tests experiments encumbered with limited budget finite computational resources. Moreover, practical scenarios in the industry, impose restrictions, based on logistics and preference, manner which can be conducted. For example, material supply may enable only a handful of single-shot case models one face significant wait-time shared In such scenarios, usually resorts to performing allows for maximizing one’s state-of-knowledge while satisfying above-mentioned constraints. Sequential design (SDOE) is popular suite have yielded promising results recent years across different engineering problems. A common strategy leverages Bayesian formalism SDOE, works best one-step-ahead myopic scenario selecting single experiment at each step sequence experiments. this work, we aim extend SDOE strategy, query computer code batch inputs. To end, leverage deep reinforcement learning (RL)-based policy gradient methods, propose batches queries selected taking into account entire hand. The algorithm retains sequential nature, inherent incorporating elements reward task from domain RL. unique capability proposed methodology its ability applied multiple tasks, optimization function, once trained. We demonstrate performance synthetic problem challenging high-dimensional problem.