Green Computing in Heterogeneous Internet of Things: Optimizing Energy Allocation Using SARSA-based Reinforcement Learning

Green computing has been emerged as a promising paradigm to harvest energy from renewable resources to reduce use of conventional energy source in Internet of Things. Energy harvested battery-operated computation chargers are deployed in network to increase energy efficiency and network lifetime by recharging the nodes. The goal of paper is to allocate optimal power and optimal channel distribution to each node in network, subject to optimize the energy efficiency of network. As, wireless channel and energy harvested from environment are stochastic in nature. So, we propose a model-free 'on policy' based reinforcement learning (RL) approach to learn transition probability of continuous state space to get optimal reward which is unknown to network. To solve the formulated problem, an RL agent uses state-action-reward-state-action (SARSA) algorithm with linear function approximation to optimize the reward. Further, the proposed SARSA algorithm is compared with Q-Learning and Myopic algorithm to analyze the network energy efficiency.