Аннотация
Algorithm selection is a general purpose optimization based on meta-learning. It uses meta-features to select algorithms on a case by case basis for the most accurate result. In real world problems such as image understanding or image segmentation the meta-features are difficult to obtain due to high diversity and lack of structure in the input data. Therefore algorithm selection in real world problems has been traditionally using image features that are also used for the task processing. Such features however might are not well suited for algorithm selection due to their task specific purpose. In order to solve this problem we propose an improved verification based approach to algorithm selection. First, we build a verification table indicating each algorithms ability to generate content that would satisfy verification for each verification attribute. Using these statistics as basis, we create a multi-algorithm verification map (MAVM). This map is built as a statistical co-occurrences of individual algorithm's ability to satisfy relational verification process of different attributes. The MAVM is reformatted so that for each relational attribute we obtain a co-occurrence of success of each algorithm. The resulting table then used for selection based on the verification result given an algorithm. The advantages of the proposed method are lesser computational cost, possibility to determine ideal combination of algorithms on a case by case basis and provides a basis for algorithm design.