Optimizing Energy Consumption in a Real-Time System Using Artificial Intelligence

Abstract: In energy-efficient real-time embedded system design, the objective is to reduce energy consumption while meeting the tasks' timing requirements. Real-time Dynamic Voltage and Frequency Scaling (DVFS) methods aim at achieving this by scaling the frequency at which a single processor or multiple processors in the system operate, but they often assume that the tasks' deadlines are known and their arrival times are regular. In this thesis project, artificial intelligence (AI) methods are explored to scale the frequency of a multiprocessor real-time system where the deadlines are too impractical to compute or defined too loosely and the arrival times are irregular. Among the studied AI methods are the reinforcement learning techniques of Q-learning and deep Q-learning, and metaheuristic genetic algorithms. The frequency scaling solutions acquired through these approaches are able to improve power savings while meeting the tasks' timing requirements. As the highest workload among processor workloads determines the frequency that the group of processors is scaled to, further, a deep Q-learning approach tobalance the loads is proposed. The allocation solutions acquired through this approach, thus far, do not significantly improve the power savings and further investigation of the learning parameters is suggested.