Efficient multi-hop data dissemination is a crucial building block to enable mature wireless sensor network (WSN) applications. Exploiting machine learning for these routing problems has received increasing attention in recent years due to its flexibility and localized mechanisms. However, with such an approach the resulting protocols often have additional memory and processing time requirements. Nevertheless, these requirements \emph {are} within the reach of today`s WSN hardware, however few substantial tests have been performed to clearly demonstrate this. This paper evaluates and discusses the results and experiences gained from implementing our reinforcement learning based multicast routing protocol (FROMS) in a testbed of ScatterWeb nodes. A comparison of our results is made to a well-known WSN routing scheme, namely a multicast variation of Directed Diffusion. Our evaluation includes several minor, but practical modifications to both protocols such as transmission backoffs and the use of acknowledgments.This paper offers three main contributions. First, we demonstrate that machine learning algorithms \emph{can be} efficiently implemented on resource restricted devices and that they perform very well in multiple network scenarios. Second, we confirm the validity of simulation results obtained in a previous evaluation of FROMS, and at the same time gather delivery rates under realistic settings. Finally, we offer some general observations on properties and pitfalls of WSN implementations along with potential solutions.
An Efficient Implementation of Reinforcement Learning Based Routing on Real WSN Hardware
Murphy, Amy Lynn;
2008-01-01
Abstract
Efficient multi-hop data dissemination is a crucial building block to enable mature wireless sensor network (WSN) applications. Exploiting machine learning for these routing problems has received increasing attention in recent years due to its flexibility and localized mechanisms. However, with such an approach the resulting protocols often have additional memory and processing time requirements. Nevertheless, these requirements \emph {are} within the reach of today`s WSN hardware, however few substantial tests have been performed to clearly demonstrate this. This paper evaluates and discusses the results and experiences gained from implementing our reinforcement learning based multicast routing protocol (FROMS) in a testbed of ScatterWeb nodes. A comparison of our results is made to a well-known WSN routing scheme, namely a multicast variation of Directed Diffusion. Our evaluation includes several minor, but practical modifications to both protocols such as transmission backoffs and the use of acknowledgments.This paper offers three main contributions. First, we demonstrate that machine learning algorithms \emph{can be} efficiently implemented on resource restricted devices and that they perform very well in multiple network scenarios. Second, we confirm the validity of simulation results obtained in a previous evaluation of FROMS, and at the same time gather delivery rates under realistic settings. Finally, we offer some general observations on properties and pitfalls of WSN implementations along with potential solutions.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
