Feature selection with acquisition cost for optimizing sensor system design
Selection of variables from large sets of measurements is a common problem of data analysis and signal processing in many disciplines. In engineering and sensor technology the design of recognition systems can be optimized by judicious choice of subsets of relevant features. In particular, the effort required for signal processing and sensor registration can be considerably reduced by efficient feature selection. However, the current approaches in majority only consider the contribution of features or measurements to the classification ability of the system. The associated cost in terms of computation effort, the required electronics, and power dissipation is not explicitly in consideration. This paper proposes a multi-objective extension of feature selection including acquisition cost and employing and comparing two evolutionary optimization methods. The genetic and particle swarm algorithms and the results achieved with selected data sets will be presented. The results show, that particle swarm algorithm can select best features with lower cost and achieve more competitive results with regard to convergence time and classification accuracy than genetic algorithm.