Reinforcement learning based MPC with neural dynamical models

This paper presents an end-to-end learning approach to developing a Nonlinear Model Predictive Control (NMPC) policy, which does not require an explicit first-principles model and assumes that the system dynamics are either unknown or partially known. The paper proposes the use of available measurements to identify a nominal Recurrent Neural Network (RNN) model to capture the nonlinear dynamics, which includes constraints on the state variables and inputs. To address the issue of suboptimal control policies resulting from simply fitting the model to the data, this paper uses Reinforcement learning (RL) to tune the NMPC scheme and generate an optimal policy for the real system. The approach's novelty lies in the use of RL to overcome the limitations of the nominal RNN model and generate a more accurate control policy. The paper discusses the implementation aspects of initial state estimation for RNN models and integration of neural models in MPC. The presented method is demonstrated on a classic benchmark control problem: cascaded two tank system (CTS).