Distributed Model-Free Stochastic Optimization in Wireless Sensor Networks

Abstract

With the improvement in computer electronics in terms of processing, memory and communication capabilities, it has become possible to scatter tiny embedded devices such as sensor nodes to monitor physical phenomena with greater flexibility. A large number of sensor nodes, communicating over the wireless medium, also allows information gathering with greater accuracy than current systems. This paper presents a new stochastic technique known as Incremental Simultaneous Perturbation Approximation(ISPA) for performing optimization in wireless sensor networks. The proposed algorithm is based on a combination of gradient-based decentralized incremental (GBDI) optimization and Simultaneous Perturbation Stochastic Approximation (SPSA) techniques. The former is based on Incremental Sub-Gradient Optimization (ISGO) techniques that allow the algorithm to be performed in a distributed and collaborative manner. The latter component addresses the limitations of the GBDI component especially in real-world sensor networks. Specifically, the SPSA component is a model-free technique that finds the optimal solution without requiring a functional model such as an input-output relationship and a cost gradient. Simulation results show that the proposed ISPA approach not only achieves distributed optimization in a stochastic environment, but can also be implemented in a practical manner for resource-constrained devices.