Knowledge in Memetic Algorithms for Stock Classification

This paper introduces a framework for a knowledge-based memetic algorithm, called KBMA. The problem of stock classification is the test bed for the performance of KBMA. Domain knowledge is incorporated into the initialization and reproduction phases of evolutionary computation. In particular, the structure of financial statements is used to sort the attributes, which contributed to a faster convergence on near optimal solutions. A semantic net is used to measure the distance between parents and offspring. Two case studies were implemented, in which domain knowledge is used to constrain the reproductive operators so that the offspring is semantically dissimilar or similar to the parent. The results show that KBMA outperformed the random memetic algorithm in the former case but did not in the latter case. The interpretation of the results is that when the search algorithm is distant from its goal, making large steps as defined by the semantic knowledge is helpful to the search.