Actor-Critic Enhanced Neural Network Optimizer for Mixed-Integer Optimal Control: A Case Study on Eco-Driving Control

This paper presents an algorithm of Actor-Critic enhanced neural network optimizer, in order to improve the optimization performance of mixed integer control problems. The proposed algorithm first utilizes an neural network optimizer with a soft-argmax operator to handle the integer variables of the optimization problem in predictive horizon. Then, considering the possible local optimality of integer control inputs, the actor-critic algorithm in predictive horizon is adopted to handle only integer variables to further enhance the optimality of the solution based on the training data from the previous optimization stage. The evaluation of algorithm performance is deployed in a case study of eco-driving control problem, and the simulation results show the superiorities over the benchmark algorithm of the cost function values in predictive horizon.