Reinforcement Learning for the Control of Bacterial Growth Bioprocess

paper deals with the application of the reinforcement learning technique (from the well-known field of machine learning) for the control of a biotechnological process – the bacterial growth. The mathematical model of the bioprocess consists of a set of nonlinear differential equations. This type of systems is difficult to control because they are strongly nonlinear, with time-varying parameters and very sensitive to small changes in initial conditions. The paper presents a control strategy based on an actor-critic structure from the reinforcement learning approach that is a combination of optimal control (based on state feedback) and adaptive control. The adaptive part of the controller is a neural network used for value function approximation. The performance of the algorithm is analyzed using numerical simulations.