[论文解读] Optimal Relay Selection with Channel Probing in Wireless Sensor Networks

Kolar Purushothama Naveen, Anurag Kumar|arXiv (Cornell University)|Jul 28, 2011

Energy Efficient Wireless Sensor Networks参考文献 21被引用 3

一句话总结

本文提出了一种基于马尔可夫决策过程的无线传感器网络中继选择方案，其中节点随机唤醒并仅揭示奖励分布，需通过探测才能获知确切奖励。最优策略被证明是在受限策略类中采用基于阈值的停止规则，其性能接近无限制情况，在满足奖励约束下最小化转发延迟。

ABSTRACT

Motivated by the problem of distributed geographical packet forwarding in a wireless sensor network with sleep-wake cycling nodes, we propose a local forwarding model comprising a node that wishes to forward a packet towards a destination, and a set of next-hop relay nodes, each of which is associated with a reward that summarises the cost/benefit of forwarding the packet through that relay. The relays wake up at random times, at which instants they reveal only the probability distributions of their rewards (e.g., by revealing their locations). To determine a relay's exact reward, the forwarding node has to further probe the relay, incurring a probing cost. Thus, at each relay wake-up instant, the source, given a set of relay reward distributions, has to decide whether to stop (and forward the packet to an already probed relay), continue waiting for further relays to wake-up, or probe an unprobed relay. We formulate the problem as a Markov decision process, with the objective being to minimize the packet forwarding delay subject to a constraint on the effective reward (the difference between the total probing cost and the actual reward of the chosen relay). Our problem can be considered as a variant of the asset selling problem with partial revelation of offers. The most general class of decision policies can keep awake any or all the relays that have woken up. In this paper, we study the optimum over a restricted class of policies which, at any time, can keep only one unprobed relay awake, in addition to the best among the probed relays. We prove that the optimum stopping policy over this class is of threshold type, where the same threshold is used at each relay wake-up instant. Numerically, we find that the performance of the optimum over the restricted class is very close to that over the unrestricted class.

研究动机与目标

• 解决具有睡眠-唤醒周期的无线传感器网络中分布式地理路由包转发的挑战。
• 将中继选择建模为马尔可夫决策过程，以在尊重奖励约束的前提下最小化转发延迟。
• 建立探测成本与中继奖励之间的权衡，初始时仅揭示奖励的概率分布。
• 研究一类受限策略，即仅允许一个未探测中继在任何时刻保持激活状态，外加最佳已探测中继。
• 确定在此受限策略类中的最优停止策略，并评估其相对于无限制情况的性能。

提出的方法

• 将中继选择问题建模为马尔可夫决策过程（MDP），其中状态由已探测中继的集合以及未探测中继的当前奖励分布定义。
• 将奖励定义为实际中继奖励与累积探测成本之间的差值，并对有效奖励施加约束。
• 采用基于阈值的停止策略，即在每次唤醒时，根据期望奖励是否超过固定阈值来决定是否探测或停止。
• 将策略类限制为任意时刻仅允许一个未探测中继保持活跃，从而简化决策空间。
• 通过动态规划和值迭代论证，证明在此受限类中，最优策略为阈值型。
• 通过数值比较阈值策略的延迟和奖励性能与无限制策略类的理论最优值，评估其性能。

实验结果

研究问题

• RQ1 当节点唤醒时仅部分揭示中继奖励信息，中继选择的最优停止策略是什么？
• RQ2 仅允许一个未探测中继保持激活的受限策略类的性能与无限制策略类相比如何？
• RQ3 基于阈值的策略能否在奖励约束下实现接近最优的转发延迟最小化？
• RQ4 探测成本对中继选择中延迟与有效奖励之间权衡的影响是什么？
• RQ5 奖励分布的结构如何影响停止规则的最优性与阈值大小？

主要发现

• 在受限策略类中，最优停止策略为阈值型，且在每个中继唤醒时刻均采用相同的阈值。
• 数值结果表明，受限策略类中最优策略的性能与无限制策略类中的最优策略非常接近。
• 阈值策略有效平衡了探测成本与奖励增益，在有效奖励约束下最小化了期望转发延迟。
• 该模型捕捉了在动态、部分可观测环境中探索（探测新中继）与利用（选择最佳已探测中继）之间的权衡。
• 采用固定阈值简化了实现，同时保持了接近最优的性能，使其适用于资源受限的传感器网络。
• 该问题被建模为部分信息揭示的资产出售问题变体，将现有理论扩展至无线网络场景。