Decentralized Convex Optimization for Wireless Sensor Networks

Many real-world applications arising in domains such as large-scale machine learning, wired and wireless networks can be formulated as distributed linear least-squares over a large network. These problems often have their data naturally distributed. For instance applications such as seismic imaging, smart grid have the sensors geographically distributed and the current algorithms to analyze these data rely on centralized approach. The data is either gathered manually, or relayed by expensive broadband stations, and then processed at a base station. This approach is time-consuming (weeks to months) and hazardous as the task involves manual data gathering in extreme conditions. To obtain the solution in real-time, we require decentralized algorithms that do not rely on a fusion center, cluster heads, or multi-hop communication. In this thesis, we propose several decentralized least squares optimization algorithm that are suitable for performing real-time seismic imaging in a sensor network. The algorithms are evaluated and tested using both synthetic and real-data traces. The results validate that our distributed algorithm is able to obtain a satisfactory image similar to centralized computation under constraints of network resources, while distributing the computational burden to sensor nodes. INDEX WORDS: Decentralized Optimization, Machine Learning, Seismic Tomography TITLE: DECENTRALIZED CONVEX OPTIMIZATION FOR WIRELESS SENSOR NETWORKS