A Parametric Bayesian Optimization Framework for Batch Dynamical Systems
We present a Bayesian Optimization (BO) framework for optimizing the performance of batch dynamical systems. A key distinctive aspect of this framework is that it uses a parametric machine learning model (a recurrent neural network - RNN) to learn the system dynamics directly from data. The use of a parametric model provides more flexibility to capture complex dynamics and to propose batches of experiments, compared to traditional BO frameworks based on non-parametric Gaussian process (GP) models. However, the use of parametric models introduces challenges in deriving and computing an information measure that can be embedded in the BO acquisition function. The proposed framework uses the expected information gain (EIG) as information measure; we argue that this enables more scalable computations compared to the use of the Fisher Information (FI) measure used in classical design of experiments. We provide a bioreactor case study to illustrate the behavior and benefits of the proposed framework.
