Reinforcement Learning for Improving Gene Identification Accuracy by Combination of Gene-Finding Programs

Due to the explosive and growing size of the genome database, the discovery of gene has become one of the most computationally intensive tasks in bioinformatics. Many such systems have been developed to find genes; however, there is still some room to improve the prediction accuracy. This paper proposes a reinforcement learning model for a combination of gene predictions from existing gene-finding programs. The model learns the optimal policy for accepting the best predictions. The fitness of a policy is reinforced if the selected prediction at a nucleotide site correctly corresponds to the true annotation. The model searches for the optimal policy which maximizes the expected prediction accuracy over all nucleotide sites in the sequences. The experimental results demonstrate that the proposed model yields higher prediction accuracy than that obtained by the single best program.