Artificial Intelligence assisted music composition has gained popularity during the last decade, but still faces problems and difficulties. This paper approaches 4-part harmonization problem in the context of evolutionary computation, a topic that has been discussed for more than forty years but still represents an open challenge. This research proposes local search to improve directed mutation and guide the algorithm towards qualitatively better solutions. Human learning and Evolutionary Machine Teaching are also used: data from music conservatory students are exploited to guide the algorithm. The results show that local search significantly improves the quality of the mutation performed. Moreover, a series of longer runs are able to find free error scores.