An Improved Q-Learning for System Power Optimization with Temperature, Performance and Energy Constraint Modeling

Abstract

Power management of embedded systems based on machine learning have drawn more and more attention. High-level software power management and optimization have gradually become important technologies for controlling the computer system power dissipation. In paper, we have employed an improved power optimization management technique which employ Q-learning algorithm based on temperature, performance and energy. The improved Q-learning has been employed to control the uncertain states of the running system and can effectively make decisions to select a rational policy with multiple parameter constraints. As running hardware and application data can be effectively collected and modeled, the power management framework can easily explore an ideal policy by value function of Q-learning algorithm.