A framework was proposed to find optimal maintenance policies in a road network. The framework included: identifying factors that contribute to policy-making; clustering the network based on these factors; identifying criteria that impact optimal policies; and determining optimal policies and periods using these criteria. To test the framework applicability, it was applied to Iran roads network step-by-step. First, road functional class and climate were identified as factors that contribute to policy-making. Second, the network was clustered into six sub-networks based on two road functional classes and three climates. Third; maintenance cost, network condition, and maintenance period were identified as criteria that impact optimal policies. Fourth, 96 maintenance policies were applied to each sub-network considering two-year, four-year, six-year, and twelve-year maintenance periods. To quantify policies cost, seven machine-learning algorithms including Gradient Boosting Regression, Lasso, Ridge, Random Forest Regression, Elastic Net, Neural Network, and Multiple Linear Regression were tested. Using the coefficient of determination (R2) as the accuracy metric, it was found that in all sub-networks the Gradient Boosting Regression had the highest accuracy on testing set (greater than 90%) while that of other algorithms was between 50% and 90%. Sub-networks condition was modeled using the Markov Chain model and was measured by the average Pavement Condition Index (PCI). Having policies cost and sub-networks PCI, the optimal policy was selected using the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS). It was concluded that in all sub-networks, the four-year maintenance period was optimal. Roads in warm zone demanded the most intense policies, followed by those in cold and humid zones. The same applied to arterial roads followed by local ones.