Online learning of timeout policies for dynamic power management
ACM Transactions on Embedded Computing Systems (TECS)
| Hi-index | 0.00 |
Dynamic Power Management (DPM) refers to a set of strategies that achieves efficient power consumption by selectively turning off (or reducing the performance of) a system components when they are idle or are serving light workloads. This paper presents a Reinforcement Learning (RL) based DPM technique for a portable, multi-camera traffic monitoring system. We target the computing hardware of the sensing platform which is the major contributor to the entire power consumption. The RL technique used for the DPM of the sensing platform uses a model-free learning algorithm that does not require a priori model of the system. In addition, a robust workload estimator based on an online, Multi-Layer Artificial Neural Network (ML-ANN) is incorporated to the learning algorithm to provide partial information about the workload and to take better decisions according to the changing workload. Based on the estimated workload and a selected power-latency tradeoff parameter, the algorithm learns to use optimal time-out values in sleep and idle modes of the computing hardware. Our results show that the learning algorithm learns an optimal DPM policy for the non-stationary workload, while significantly reducing the power consumption and keeping the system response to a desired level.