Evolutionary Kernel Learning

A Evolutionary Kernel Learning is a machine learning task that uses an evolutionary algorithm in the designing of a kernel function.

See: Kernel-based Learning Algorithm, Feature Space, Multi-Objective Optimization.

References

2011

(Igel, 2011) ⇒ Christian Igel. (2011). “Evolutionary Kernel Learning.” In: (Sammut & Webb, 2011) p.369
- QUOTE: Definition - Evolutionary kernel learning stands for using evolutionary algorithms to design the kernel function for a kernel method.
  Motivation and Background
  In kernel-based learning algorithms, the kernel function implicitly defines the feature space in which the algorithm operates. The kernel determines the scalar product and thereby the metric in the feature space. Choosing the right kernel function is crucial for the training accuracy and generalization performance of the learning machine. The choice may also influence the runtime and storage complexity during and after training.
  The kernel is usually not adapted by the kernel method itself; choosing it is a model selection problem. In practice, the kernel function is selected from an a priori fixed class. When a parameterized family of kernel functions is considered, kernel adaptation reduces to finding appropriate parameters. The most frequently used method to determine these values is grid search. In simple grid search, the parameters are varied with a fixed step-size through a range of values, and the performance of each combination is measured. Because of its computational complexity, grid search is only suitable for the adjustment of a few parameters. Furthermore, the choice of the discretization of the search space may be crucial. Gradient-based approaches are perhaps the most elaborate techniques for adapting real-valued kernel parameters, see the articles by Chapelle et al. (2002) and Glasmachers and Igel (2005) and references therein. To use these methods, however, the class of kernel functions must have a differentiable structure. Furthermore, score functions for assessing the parameter performance that are not differentiable and/or piecewise constant may cause problems. Evolutionary kernel learning does not suffer from these limitations. Additionally, it allows for multi-objective optimization (MOO) to address several kernel design criteria.