A Noise-Insensitive Reinforcement Learning Control for a Nonlinear Bioreactor

The capacity of the Model-Free Learning Control (MFLC) system is studied in the presence of noise signals for a highly nonlinear bioreactor. The bioreactor is a Multiple-Input Multiple-Output (MIMO) system based in a biochemical and physical model. A methodology is provided for designing the MFLC scheme and optimally selecting their elements and parameters. The scheme is tested for 3 different inputs (step, sinusoidal, and ramp) and it includes exogenous disturbances in the state variables. Noise signals are added to the output variables to model errors in sensor readout and uncertainties of the plant. The results show that the MFLC systems is very robust and insensitive to the presence of noise signal even when it is very high (50%). The presented methodology could be applied to control any nonlinear industrial process.