In a recent work, a learning procedure for relaxation labeling algorithms has been introduced which involves minimizing a certain cost function with classical gradient methods. Although the results obtained so far are extremely promising, the gradient-based learning algorithm suffers from some inherent drawbacks that could prevent its application to real-world problems of practical interest. Essentially, these include the inability to escape from local minima and its computational complexity. In this paper, we propose using genetic algorithms to solve the relaxation labeling learning problem in an attempt to overcome the difficulties with the gradient algorithm. Experiments are presented which demonstrate the superiority of the proposed approach both in terms of quality of solutions and robustness.