Zpětnovazební učení pro řízení optimalizovaných vestavěných systémů

Abstract

The aim of this thesis is to optimize the energy use of an independent embedded system. For long term measurements of air pollution, it is necessary to place the device in a remote location, where it is not possible to supply energy demands from power grid. For this purpose a reinforcement learning algorithm has been choosen. Which is able to learn the desired behaviour based on the prescribed policy. The thesis describes the principles of reinforcement learning and energy independent embbeded systems. Testing is done in simulation enviroment and its results are compared with time based control. For simulation, meteorogical data was used over the course of four years. The created algorithm exhibits greater robustness, minimizes supercapacitor overcharging and prevents any power outages. Based on the simulation results controlling algorithm is capable of long term control and can be used on target device.