Probabilistic Modeling for Optimization of Bioreactors using Reinforcement Learning with Active Inference

Full title: 

Probabilistic Modeling for Optimization of Bioreactors using Reinforcement Learning with Active Inference

Abstract:

The open-ended complexity of abiotic conditions in bioreactors means that it is generally infeasible to model its dynamic behaviour comprehensively. Learning optimization oriented probabilistic models encoding a parsimonious representation is far more efficient for bioprocess development and optimization. In this work, active inference is integrated with reinforcement learning to demonstrate that useful probabilistic models for bioreactor optimization can be learned by balancing optimization-oriented and information-seeking objectives. The baker’s yeast bioprocess is used as a case study. For online Bayesian update of model parameter distributions, simulation results demonstrate that highly informative data can be sampled by minimizing the variational free energy of the expected future. The resulting probabilistic model is thus biased towards bioreactor optimization.