Evolving machine learning and deep learning models using evolutionary algorithms

Xie, Hailun (2020) Evolving machine learning and deep learning models using evolutionary algorithms. Doctoral thesis, Northumbria University.

Preview

Text (Doctoral Thesis) xie.hailun_phd_16042740.pdf - Submitted Version Download (2MB) | Preview

Abstract

Despite the great success in data mining, machine learning and deep learning models are yet subject to material obstacles when tackling real-life challenges, such as feature selection, initialization sensitivity, as well as hyperparameter optimization. The prevalence of these obstacles has severely constrained conventional machine learning and deep learning methods from fulfilling their potentials. In this research, three evolving machine learning and one evolving deep learning models are proposed to eliminate above bottlenecks, i.e. improving model initialization, enhancing feature representation, as well as optimizing model configuration, respectively, through hybridization between the advanced evolutionary algorithms and the conventional ML and DL methods.

Specifically, two Firefly Algorithm based evolutionary clustering models are proposed to optimize cluster centroids in K-means and overcome initialization sensitivity as well as local stagnation. Secondly, a Particle Swarm Optimization based evolving feature selection model is developed for automatic identification of the most effective feature subset and reduction of feature dimensionality for tackling classification problems. Lastly, a Grey Wolf Optimizer based evolving Convolutional Neural Network-Long Short-Term Memory method is devised for automatic generation of the optimal topological and learning configurations for Convolutional Neural Network-Long Short-Term Memory networks to undertake multivariate time series prediction problems.

Moreover, a variety of tailored search strategies are proposed to eliminate the intrinsic limitations embedded in the search mechanisms of the three employed evolutionary algorithms, i.e. the dictation of the global best signal in Particle Swarm Optimization, the constraint of the diagonal movement in Firefly Algorithm, as well as the acute contraction of search territory in Grey Wolf Optimizer, respectively. The remedy strategies include the diversification of guiding signals, the adaptive nonlinear search parameters, the hybrid position updating mechanisms, as well as the enhancement of population leaders. As such, the enhanced Particle Swarm Optimization, Firefly Algorithm, and Grey Wolf Optimizer variants are more likely to attain global optimality on complex search landscapes embedded in data mining problems, owing to the elevated search diversity as well as the achievement of advanced trade-offs between exploration and exploitation.

Actions (login required)

View Item

Downloads