Students are required to choose courses they are interested in for the coming semester. Due to restrictions, including lack of sufficient resources and overheads of running several courses, some universities might not offer all of a student's desirable courses. Universities must know every student's demands for every course prior to each semester for optimal course scheduling. This research examines the problems associated with course selection in the context of e-learning. This study is focused on identifying the potential factors that affect student satisfaction concerning the online courses they select, modeling student course selection behavior and fitting a function to the training data using neural network approach, and applying the obtained function to predict the final number registrations in every course after the drop and add period. The experimental sample came from 714 online graduate courses in 16 academic terms from 2005 to 2012. Findings disclosed high prediction accuracy based on the experimental data and exhibited that the proposed model outperforms three well-known machine learning techniques and two previous, naive approaches significantly. This contribution finally ends with an analysis and interpretation of results, and presentation of some suggestions and recommendations for enthusiastic educational institutes regarding how to choose the best strategy and configuration to expand and also adapt the introduced system to their specific needs. (Contains 7 tables and 3 figures.)