Stochastic switching model and policy optimization online for dynamic power management

A reinforcement learning based online optimization algorithm is presented for dynamic power management with unknown system parameters. First an event-driven stochastic switching model is introduced to formulate dynamic power management problem as a constrained policy optimization problem. Then by utilizing the features of this model an online optimization algorithm that combines policy gradient estimation and stochastic approximation is derived. The stochastic switching model captures the power-managed system behaves accurately. The optimization algorithm is adaptive, and can achieve global optimum with less computational cost. Simulation results demonstrate the effectiveness of the proposed approach.