Optimization of decentralized quantizers in rate constrained data fusion systems

A common model for distributed detection systems is that of several separated sensors each of which measures some observable, quantizes it, and communicates to a fusion center the quantized observation. The fusion center collects the quantized observations and takes the decision. This paper deals with the design of the quantizers when the channels between the sensors and the fusion center are subject to capacity constraints. The system of interest allows soft nonbreakpoint quantizers and nonindependent observations. The authors investigate two optimization techniques. The first technique finds a local minimum of the average misclassification risk by alternate optimization. The computational cost of the technique is exponential in the sum of the rates of the quantizers. To overcome this difficulty a second technique based on a neural approach is presented, where a local minimum of the average misclassification risk is found by a stochastic approximation method