Combinatorial Feature Optimization using Multi-objective Evolutionary Algorithms applied to a Biological Warfare Classification Problem.

Abstract

Biological weapons is the aggressive use of organisms or toxins, also known as biological warfare agents. These weapons are invisible, odorless, tasteless and can be spread without a sound, making it difficult to detect an attack. Early warning systems based on environmental standoff detection of biological warfare agents using lidar technology require real-time signal processing, challenging the systems efficiency in terms of both computational complexity and classification accuracy. Hence, optimal feature extraction and feature selection is essential in forming a stable and efficient classification system. Optimization of signal processing with a high degree of freedom, meaning the possibilities of processing the signals are relatively unrestricted, implies high-dimensional solutions and consequently a large search space. Moreover, there is no linearity between the selection variable space of the solutions and the objective function. Thus, many classical optimization methods will be unsuitable for the task. The objective of this thesis has been employing genetic algorithms in the search for optimal features for classification of biological warfare agents. The flexibility of evolutionary algorithms enables simultaneous optimization of more than one objective function, making it possible to optimize both classification accuracy and computational complexity combinatorially. The algorithms outperform benchmark methods like Support Vector Machines, Fisher Linear Discriminant, Principal Component Analysis, Sequential Forward Feature Selection and Random Search, with significantly improved classification accuracy compared to the best classical method. The results also give valuable information related to the design of instruments and detection systems, giving new insight to signal processing.