Tournament Feature Selection with Directed Mutations

A tournament searching method with new mutation operators for a problem of the feature subset selection is presented. The probability of the bit mutation in a classical approach is fixed. In the proposed methods this probability is dependent on the history of the searching process. Bit position whose mutation from 0 to 1 (from 1 to 0) improved the evaluation of the solution in early iterations, are mutated more frequently from 0 to 1 (from 1 to 0). The roulette wheel method and the tournament method are used to select the bits for the mutation according to the adaptive probability. The algorithms were tested on several tasks of the feature selection in the supervised learning. The experiments showed the faster convergence of the algorithm with directed mutations in relation to the classical mutation.