2020/11/26 by Eivind Bøhn, Bøhn, Eivind, Sébastien Gros +5
Computer Science · Engineering · #Advanced Control Systems Optimization #FOS: Computer and information sciences #FOS: Electrical engineering #Fuel Cells and Related Materials #Machine Learning (cs.LG) #Reinforcement Learning in Robotics #Systems and Control (eess.SY) #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.2011.13365
openalex publication_date 2020/11/26 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
In control applications there is often a compromise that needs to be made\nwith regards to the complexity and performance of the controller and the\ncomputational resources that are available. For instance, the typical hardware\nplatform in embedded control applications is a microcontroller with limited\nmemory and processing power, and for battery powered applications the control\nsystem can account for a significant portion of the energy consumption. We\npropose a controller architecture in which the computational cost is explicitly\noptimized along with the control objective. This is achieved by a three-part\narchitecture where a high-level, computationally expensive controller generates\nplans, which a computationally simpler controller executes by compensating for\nprediction errors, while a recomputation policy decides when the plan should be\nrecomputed. In this paper, we employ model predictive control (MPC) as the\nhigh-level plan-generating controller, a linear state feedback controller as\nthe simpler compensating controller, and reinforcement learning (RL) to learn\nthe recomputation policy. Simulation results for two examples showcase the\narchitecture's ability to improve upon the MPC approach and find reasonable\ncompromises weighing the performance on the control objective and the\ncomputational resources expended.\n