A Novel Fully Evolved Kernel Method for Feature Computation from Multisensor Signal Using Evolutionary Algorithms

Abstract

The design of intelligent sensor systems requires sophisticated methods from conventional signal processing and computational intelligence. Currently, a significant part of the overall system architecture still has to be manually elaborated in a tedious and time consuming process by an experienced designer for each new application or modification. Clearly, an automatic method for auto-configuration of sensor systems would be salient. In this paper, we contribute to the optimization of the feature computation step in the overall system design, investigating Gaussian kernel methods. Our goal is to improve the kernel method of feature computation with consideration on including the adjustable magnitude parameter for Gaussian kernels or fully evolved Gaussian kernels, which are inspired by feature weighting concepts and are similar to RBF like neural network with correlation based kernel layer and linear combiner output layer. We compare this improved method with previous kernel methods using weighting method of multiobjective evolutionary optimization, i.e., genetic algorithms. In addition to the straightforward feature space from the optimized kernel layer, we complement the kernel layer by linear combiner layer, with weights computed by traditional IDA (linear discriminant analysis) in the loop of the optimization. In our experiments, we applied gas sensor benchmark data and the results showed that our novel method can achieve competitive or even better recognition accuracies and effectively reduce the computational complexity as well.