The estimation of distribution algorithms (EDA) and the Learnable Evolution Model (LEM) methods are both good examples of combining machine learning and population based evolutionary search.. The basic point is to by extracting the high-level models of the variable relationships, we could make better decision when generating the new exploration individuals.
How could we apply this ideas to Genetic Programming? Sastry Kumar did a try but that kind of fixed full-tree model is of limited value. A better method is needed.