Sintonia de Controladores PI baseada em Augmented Random Search: Estudo de Caso do Processo CSTR

The ARS algorithm is a reinforcement learning method that seeks to map the best actions to the process operating conditions. Initially, the algorithm has no instructions and no knowledge of the process dynamics. Thus, this technique seeks to learn while interacting with the process and the search for the best actions is guided only by a numerical reward signal. The continuous stirring tank (CSTR) process, simulated in Python, is used as an interaction environment for the ARS algorithm. The main objective of this work is to apply the ARS algorithm to tune a PI controller of the CSTR process. The states are the setpoint values applied to the process before and after its variation. Actions are the PI controller parameters for each reference set (states). The reward was defined as the inverse of the sum of the error module. The tunings proposed by the ARS are 8.3% (same operating point) better than the tuning benchmark chosen for comparison.