Abstract
Biopharmaceutical manufacturing faces critical challenges, including complexity, high variability, lengthy lead time, and limited historical data and domain knowledge of the underlying system stochastic process. To address these challenges, we propose a green simulation assisted model-based reinforcement learning (GS-RL) method where the bioprocess model risk is quantified by the posterior distribution given observed data and all simulation outputs in the learning process are being efficiently recycled and reused. The main benefit of the proposed method is high computational efficiency, as it simultaneously guides learning and dynamic decision making. The green simulation likelihood ratio metamodel reuses simulation outputs from previous iterations in a stochastic search algorithm for the optimal policy and the outputs from previous experiments. As such, the quality of gradient estimation is improved and the search for the optimal policy converges faster. Our numerical studies show promising results.