Optimal control of networked dynamic systems via input convex graph recurrent neural network

Abstract

To control the dynamic system with model-based approach, the model should represent the dynamic of the system and at the same time, the model should not be computational intractable on control optimization problem. In this paper, we offer a balanced model between called input convex graph recurrent neural network to harmonize model expressivity and optimization solvability. Input convex graph recurrent neural network is a graph recurrent neural network which output of the function is a convex function with respect to input of the function. The graph recurrent neural network structure in the proposed model can represent the interaction between components of system and the dynamic of system efficiently. In addition, input convexity of the proposed model guarantee to find an optimal solution when the agents find the action sequence using the dynamic model. We apply the proposed dynamic model to control the furnace system in the real world with model predictive control. The experimental shows that the proposed model shows notable performance than baseline models.