Online Optimization For Dynamic Power Management

With the rapid development of Internet and network communication, electronic devices like smartphones, laptops have been integrated deeply into our daily life. We take the smart phone as an example, with its function more and richer, the display screen is becoming bigger and bigger, the degree of intelligence is higher and higher. However, the current power management strategies usually adopt the empirical strategy. Therefore, this dissertation proposed online optimization method of power management based on the Time-Index Semi-Markov decision process(TI-SMDP) model with randomized strategy, which purpose is to reduce energy consumption of equipment and growth equipment standby time.The dynamic power management problem will be abstracted into an optimization problem with constraint, and then, by using the method based on spherical coordinates, the optimization problem of the constrained optimization problem becomes an optimization problem with the introduction of parameters. Next, we will transform a constrained optimization problem into an unconstrained optimization by means of Hestenes-Powell augmented Lagrange function method, and by using the transformed solution of the optimization problem, we can get the optimal solution to the original problem. For the unconstrained optimization problem, we use the gradient estimation method based on the single sample path to get the optimal parameters of the parameters, and the random strategy ca n be used to make the equipment transfer to the appropriate state to achieve maximum energy saving. We call this algorithm L-C algorithm.Based on L-C algorithm, we set up a region of the device’s idle state to prohibit the equipment to sleep. We propose a method to manage the online dynamic power supply management with no region. The method can only determine the state transfer of the equipment in the idle state when the equipment is in the idle state. When we select the appropriate device to prohibit the region, it can get better results than the L-C algorithm. In the two algorithms, we use the parametric method based on spherical coordinates, which not only speeds up the converge nce rate of optimization results, but also reduced the number of algorithm parameters, thereby reducing the storage space required by the algorithm; We adopt the Hestenes-Powell augmented Lagrange method, simplifying the optimization problem, and the convergence process is more smooth; We estimate the parameters by using the gradient estimation method based on the single sample path, which can reduce the computational complexity of the algorithm.