Methods, systems, and computer program products are provided for supervised power management between a primary platform and a secondary platform. Communication between a primary platform and a secondary platform is established. An application running on the secondary platform is captured. Input features and output measures are collected to build a training set for the application, wherein the input features are collected through direct measurement and the output measures reflect characteristics of the application. Based on the training set, power consumption of the secondary platform with an expected performance level is predicted for a new application running on the secondary platform. Accordingly, an optimal power management policy is derived that minimizes the total power consumption of the primary and secondary platforms.