Meta-learning via search combined with parameter optimization

Framework for Similarity-Based Methods (SBMs) allows to create many algorithms that differ in important aspects. Although no single learning algorithm may outperform other algorithms on all data an almost optimal algorithm may be found within the SBM framework. To avoid tedious experimentation a meta-learning search procedure in the space of all possible algorithms is used to build new algorithms. Each new algorithm is generated by applying admissible extensions to the existing algorithms and the most promising are retained and extended further. Training is performed using parameter optimization techniques. Preliminary tests of this approach are very encouraging.