Most machine learning models has hyperparameters which indirectly
influences the quality of predictions. Common approaches for hyperparameter
optimization are grid search, randomized search or an approach
using evolutionary algorithms.
In this bachelor thesis I explored how the use of multiparent recombination
in evolutionary algorithms for this problem influences the quality
of the found solution. For this purpose an evolutionary algorithm for
hyperparameter optimization was developed.
In testing the algorithm outperformed grid search and randomized search.
Among the multiparent recombination algorithms the diagonal crossover
performed somewhat better than the k-point crossover. However, in
general we could not claim that the multiparent approach is superior to
the traditional one.
|
