Ricerc@Sapienza

Superior performance in wireless sensor networks is obtained by taking key protocol decisions based on the outcome of local learning-based computations, informing nodes on past and expected availability of resources. This paper investigates the impact on protocol performance of local computational requirements of learning techniques. We consider a recent routing solution, named WHARP, which makes decentralized and proactive decisions based on a Markov Decision Process (MDP) that takes into account key parameters of wireless green networks, including energy harvesting capabilities, and wakeup radio technology. We show that in these scenarios solving the MDP incurs energy expenditures by far superior to that required by wireless communication, even at very high data traffic. In order to maintain the performance advantages of the learning-based protocol machinery, we propose a heuristic solution that closely approximates the MDP trading off optimality for considerably lighter computational requirements. We compare the performance of the heuristic-based WHARP (called W-HEU) to that of the MDP-based WHARP that uses the standard Backward Value Iteration (W-BVI) through GreenCastalia-based simulations with real computational energy measurements. Our results show that W-HEU outperforms W-BVI on key metrics such as energy consumption and packet delivery ratio, making up for the lost optimality of BVI through the remarkable energy savings of its lighter computational requirements.