This paper addresses the problem of robustly estimating the epipolar geometry by employing a new technique based on messy genetic algorithms, which uses each gene to stand for a pair of correspondences, and takes every chromosome as a minimum subset for epipolar geometry estimation. The method would eventually converge to a nearly optimal solution and is relatively unaffected by outliers. Experiments with both synthetic data and real images show that our method is more robust and accurate than other typical methods because it can efficiently detect and delete the bad corresponding points, which include both bad locations and false matches.