Summary: This paper proposes an extension of Genetic Programming (GP) for agent learning. In order to improve the performance, GP with control node (GPCN) and its three kinds of modification have been proposed. In GPCN, an individual consists of several trees which have the number P of executions. The three kinds of modification are the conditional probability, the cross-cultural island model, and updating the value P. This paper proposes a new method that combines the conditional probability with the cross-cultural island method. Experiments are conducted to show the performance in the garbage collection problem and Santa Fe Trail problem.