Machine learning in wireless sensor networks

Mike Colagrosso

Dept. of Math. and Computer Sciences, Colorado School of Mines

Wed May 24 14:10:00 NZST 2006 in Room 031, MSCS

Abstract

Because wireless sensor networks comprise low-power, resource-constrained devices, my students and I apply machine learning in the network with the goal of saving energy and optimizing the usefulness of the network and the data it collects. We consider wireless sensor networks at three levels of abstraction, the micro-sensor, sensor, and macro-sensor level, and I will present machine learning methods at each level.

At the micro-sensor level, we have developed BoostMAC, a new radio communication protocol that is suitable for bursty network traffic, such as the traffic pattern encountered in a monitoring application. In BoostMAC, we apply supervised learning to optimize only a small component of a sensor: the communication schedule its radio. Because the radio of a sensor consumes so much of its energy, we show that automatically adapting a sensor's communication saves energy.

At the sensor level, the Learning to Live project creates a reinforcement learning agent on every sensor. The agent exists in an abstract set of states, and it has many actions that can take, including turning on and off its sensors and radio and forwarding data from neighboring sensors. By rewarding sensors when the wireless sensor network meets its quality of service demands, we train each sensor to learn an optimal policy, which is a function that maps states to actions. By learning many policies in simulation and analyzing them, we can define different roles for sensors in the network, such as gateway node, edge node, and sensing cluster. If sensors can choose role when they are deployed (and naturally change roles over time), then the lifetime and coverage of the network will be increased.

At the macro-sensor level, we study emergent behavior of the wireless sensor network. We have developed a localization method, titled LOST (localization of sensors in Tenet). Tenet is simply a tiered-network architecture, and at the macro-sensor level, we use this architecture for naming and routing. On top of it, our method creates a network-wide coordinate system for the sensors built using only local, pairwise distance estimates. The distance estimates can be noisy in our method, and the estimates can also be a hybrid of, say, signal strength, time of flight difference, and hop count. We collect all the distance estimates and use the Locally Linear Embedding dimensionality reduction method to create our local coordinate system, a technique from unsupervised learning. If the network has four known locations in 3D (or three known locations in 2D), then we can transform our local coordinate system into estimates of the true location of every sensor in the network.