Approximating optimal controls with recurrent neural networks for automotive systems

We discuss ways of approximating optimal control sequences by neural networks in complex automotive systems. Our systems include engines and exhaust aftertreatment modules. Our goal is to minimize both emissions and fuel consumption. As our baseline, we use open-loop solutions based on dynamic programming (DP). First, we consider the case of a recurrent neural network trained directly on a DP solution, whether discrete- or continuous-valued. We show that close approximations of DP solutions are possible with neural networks, acting as feedback controller. Second, we discuss an iterative procedure which allows neurocontrollers to approximate DP solutions indirectly. Discussion is supported by high-fidelity simulation results.