Pairwise Comparison of Hypotheses in Evolutionary Learning

This paper investigates the use of evolutionary algorithms for the search of hypothesis space in machine learning tasks. As opposed to the common scalar evaluation function imposing a complete order onto the hypothesis space, we propose genetic search incorporating pairwise comparison of hypotheses. Particularly, we allow incomparability of hypotheses, what implies a partial order in the hypothesis space. We claim that such an extension protects the ‘interesting’ hypotheses from being discarded in the search process, and thus increases the diversity of the population, allowing better exploration of the solution space. As a result it is more probable to reach hypotheses with good predictive accuracy. This supposition has been positively verified in an extensive comparative experiment of evolutionary visual learning concerning the recognition of handwritten characters.