Distributed support vector machines in sensor-actuator networks

A flat sensor network collects enormous amounts of data that can be used to train a centralized support vector machine for classification purpose. However, a flat network structure that mandates sensors exchanging data may either require large bandwidth or incur significant energy expenditure. A more efficient alternative is supported by sensor-actuator networks that rely on actuators as resource-rich intercessors to reduce energy consumption and maintain the overall task performance. We study a support vector machine (SVM) with distributive data over a sensor-actuator network. Each sensor that receives request from an actuator performs a sensing task and reports data to that actuator. Each actuator collects data, trains a local SVM through message exchanging with other actuators, and adjusts its controllable state according to its SVM result. Of our interest is the impact of actuator coordination and sensor-actuator formation on the result of distributed SVM, which is compared with a centralized SVM enabled by a flat actuator-less sensor network with one fusion site.