This paper applies a problem decomposition approach in order to solve hard Frequency Assignment Problem instances with standard meta-heuristics. The proposed technique aims to divide the initial problem into a number of easier subproblems, which can then be solved either independently or in sequence respecting the constraints between them. Finally, partial subproblems solutions are recomposed into a solution of the original problem. Our results focus on the COST-259 MI-FAP instances, for which some good assignments produced by local search meta-heuristics are widely available. However, standard implementations do not usually produce the best performance and, in particular, no good results have been previously obtained using evolutionary techniques. We show that problem decomposition can improve standard heuristics, both in terms of solution quality and runtime. Furthermore, genetic algorithms seem to benefit more from this approach, showing a higher percentage improvement, therefore reducing the gap with other local search methods.